The Library, Bay 01
The Lexicon
75 AI terms, sorted onto five shelves by where they show up, or see them all at once. Tap a book, or a note, to open it.
Human in the Loop
The namesake shelf: governance, guardrails, and keeping people in charge of AI.
Claude
Anthropic's Claude: agents, Claude Code, and the architecture behind them.
Copilot
Microsoft Copilot terminology. This shelf is growing.
Microsoft
The wider Microsoft AI stack. Books arriving.
General GenAI
Concepts that apply no matter whose AI you use.
Human in the Loop shelf
Governance & Guardrails
10 terms on this shelf.
- AI governanceThe rules for how an organization uses AI: who may use it, for what, with which data, and who is accountable when it goes wrong. Guardrails are how those rules get enforced.
- Approval gateA required human yes before an agent's action takes effect: sending, spending, publishing, deleting. The loop pauses until a person signs off.
- Audit trailThe record of what an AI system did and why: every action, tool call, and decision logged so a person can reconstruct it later. For compliance purposes, if it isn't logged, it didn't happen.
- Data sensitivityKnowing which information must never reach an AI tool (regulated personal data, client confidences, secrets) and anonymizing or withholding it before use. The redact-before-upload habit.
- GuardrailsThe technical limits that stop an AI system from doing what policy forbids: permission rules, blocked actions, filtered inputs and outputs. Policy says what; guardrails make sure.
- HallucinationWhen a model states something false with full confidence, including invented citations, numbers, or names. It isn't lying; it's filling gaps convincingly. This is the reason verification and human review exist.
- Human in the loopA person stays close enough to an AI system's work to check it, correct it, and decide when it can run on its own versus when a person must step back in. It is a design choice about where human judgment sits in a process, not a slogan.
- Human on the loopThe lighter version of human in the loop: the AI acts on its own and a person reviews afterward, sampling results and stepping in when something looks off. In the path versus watching the path.
- Least privilegeGive an AI system the smallest access it needs and nothing more.
- Red teamingDeliberately attacking your own AI system: prompting it to misbehave, leak data, or break its rules, so you find the failure before a stranger does.
Human in the Loop shelf
Claude Code
3 terms on this shelf.
- dangerously-skip-permissionsThe flag that bypasses every permission prompt. Never use it, except in strictly controlled, low-risk environments like a locked-down CI job.
- Permission rules (allow / ask / deny)Per-tool access control for Claude Code: allow silently, ask first, or deny outright, for tools like bash, read, edit, and web fetch.
- SandboxingRestricting what an agent's shell can touch on the host machine (tools like bubblewrap and socat). Containment beats trust.
Human in the Loop shelf
Prompts & Structured Output
1 term on this shelf.
Human in the Loop shelf
Context & Reliability
1 term on this shelf.
Claude shelf
Agents & Orchestration
22 terms on this shelf.
- Adaptive decompositionSplitting open-ended work (research, investigation) where findings reshape what the next subtask should be.
- AgentAn AI model running in a loop with tools: it thinks, takes an action, looks at the result, and goes again until the job is done.
- Agent SDKAnthropic's framework that runs the agent loop for you: manages subagents, ships built-in tools. The raw API makes you write the loop yourself.
- Agentic loopThe cycle your code runs around a stateless model: gather context, take action, verify results, repeat.
- API (Application Programming Interface)The doorway a program uses to talk to another program or service over the web: send a request, get a response. Claude itself is reached through one.
- Call-inspect-execute-append cycleThe full agent turn: call the API, inspect the stop_reason, execute the requested tool locally, append the result to the history, call again. Skip inspect and the agent goes silent; skip append and it repeats the same tool call forever.
- Coordinator (orchestrator)The manager agent that breaks a big job into pieces, hands them to subagents, and combines the results.
- end_turnA stop_reason value meaning "I'm done." The loop ends here, and only here.
- fork_sessionStarts a new session as a copy of an existing session's context, so parallel work can branch from the same starting point.
- HookA small script that runs automatically at a fixed moment in an agent's work. Rules that must ALWAYS happen belong in hooks. Prompts are requests; hooks are law.
- Hub and spokeThe orchestration shape where workers talk only to the manager, never to each other. Keeps the system predictable and debuggable.
- PreToolUse / PostToolUseThe two classic hook moments: right before a tool runs (gate it) and right after (enforce follow-ups, like logging every change).
- Refinement loopA pass where an agent's work is checked for coverage and gaps before the final version ships.
- SDK (Software Development Kit)A code library that wraps an API so developers do not have to hand-build every request. Anthropic ships official Python and TypeScript SDKs for Claude.
- Sequential decompositionSplitting a job into steps where each needs the previous step's output. Use when order matters.
- StatelessThe model keeps no memory between API calls. Every call starts blank; your code supplies the history each time.
- stop_reasonThe machine-readable label on every API reply saying why the model stopped. Your loop should be driven by this label, never by reading the model's prose or counting turns.
- SubagentA worker agent spawned for one piece of a job. Returns a compact summary to the coordinator, which protects the coordinator's context budget.
- Targeted retryWhen one subagent fails, retry only that one. A blanket retry re-runs steps that already succeeded, re-firing their side effects (like duplicate database writes).
- Task toolHow a coordinator in the Agent SDK/Claude Code actually spawns a subagent.
- The four surfacesThe map of the Claude ecosystem: the Claude API (one stateless HTTP endpoint plus SDKs), the Agent SDK (runs the loop), Claude Code (a terminal agent with file access), and MCP (the standard plug for outside tools).
- tool_useA stop_reason value meaning "run this tool for me and send back the result." The loop continues.
Claude shelf
Claude Code
14 terms on this shelf.
- .claude folderThe project's Claude configuration home: rules/ (always-on instructions), commands/ (reusable /name prompts), skills/ (bigger packaged abilities).
- .mcp.json vs ~/.claude.jsonWhere tool connections live: .mcp.json in the project is shared with the team; ~/.claude.json is personal. Same logic as project vs personal CLAUDE.md.
- @claude (GitHub Actions)Mentioning @claude on a GitHub issue or PR to trigger automated review or fixes through a GitHub Action.
- @importHow a CLAUDE.md pulls other files into itself, so instructions can be split up but load together.
- Claude CodeAnthropic's terminal agent. It works inside a folder or repository with real file access through tools like bash and grep.
- CLAUDE.mdThe instruction file Claude Code reads at session start. Three levels: user (~/.claude/, personal, never version-controlled), project (repo root, committed, team rules), and directory (subfolder-specific rules).
- dangerously-skip-permissionsThe flag that bypasses every permission prompt. Never use it, except in strictly controlled, low-risk environments like a locked-down CI job.
- Headless mode (-p)Runs Claude Code non-interactively with one prompt, for scripts and CI/CD. Pair with --output-format json for machine-readable results.
- Permission rules (allow / ask / deny)Per-tool access control for Claude Code: allow silently, ask first, or deny outright, for tools like bash, read, edit, and web fetch.
- Plan modeMakes Claude Code propose an approach for approval before touching anything. Right for big risky changes; overkill for one-liners.
- SandboxingRestricting what an agent's shell can touch on the host machine (tools like bubblewrap and socat). Containment beats trust.
- Session commands (resume, fork, compact, clear, rewind)Managing a session's life: re-enter an old one (resume), branch its history (fork), shrink token use (compact), wipe it (clear), roll back to an earlier turn (rewind).
- SKILL.md frontmatterThe header block of a skill file. Can set context: fork, allowed-tools, and argument-hint.
- Slash commandA reusable prompt saved as a markdown file in .claude/commands, triggered by typing /name. Its YAML frontmatter can restrict tools and hint at arguments.
Claude shelf
Context & Reliability
9 terms on this shelf.
- Context budgetTreating context like money: spend it on what matters, don't dump raw tool output into it, make subagents return summaries.
- Context windowEverything the model can see at once: the conversation, documents, tool results. Finite, and managing it is a real skill.
- Escalation triggersObjective, checkable conditions for handing off to a human: the customer asked, the request is outside policy, or no progress is being made. Never the model's detected sentiment or self-reported confidence.
- Fact block / case blockA pinned, structured block of durable facts (customer, order number, issue) re-anchored at the end of the context every turn, instead of hoping the model re-finds them on page 40.
- LLM (Large Language Model)The engine under Claude, Copilot, and their peers: a model trained on enormous amounts of text that predicts language, which turns out to be enough to draft, summarize, reason, and converse.
- Lost in the middleModels pay the most attention to the start and end of a long context. Critical facts buried in the middle get missed.
- Prompt cachingMarking static sections (system prompt, few-shot examples) with cache breakpoints so repeat calls reuse them, cutting cost up to ~90% on the cached portion.
- ScratchpadHaving a long-running agent write intermediate notes to a file instead of holding everything in context.
- TokenThe unit models actually read and write: chunks of a few characters each. Context windows, pricing, and limits are all measured in tokens, roughly three quarters of a word apiece.
Claude shelf
Prompts & Structured Output
8 terms on this shelf.
- Confidence calibration (field-level)In extraction pipelines, scoring confidence per field (99% on the date, 40% on the vendor) so low-confidence fields route to human review. Different from escalating a conversation on the model's mood, which is always wrong.
- Few-shot examplesShowing 2 to 4 examples of the output you want inside the prompt. Beats describing the format in words.
- JSON schema (as a tool)Defining the exact shape of data you need as a tool definition, so the model fills in a form instead of writing an essay you have to parse.
- Mechanically checkable rulesReplacing vague adjectives ("be careful") with categorical rules a checker could verify. Vague adjectives produce false positives.
- Message Batches APISubmit thousands of independent prompts as an offline batch: results within 24 hours at 50% of the cost. Never for anything a person is waiting on.
- Nullable fieldsMarking schema fields optional so "not present" is a legal answer. A required field with missing source data gets fabricated. Required + missing = made up.
- tool_choicePer-request control of tool use: auto (model decides), any (must use some tool), a named tool (must use that one), or none.
- Validation-retry loopCheck structured output against the schema; if invalid, send the error back for a corrected retry (usually succeeds), then fail loudly. Never silently accept bad data.
Claude shelf
Tools & MCP
7 terms on this shelf.
- MCP (Model Context Protocol)The open standard that lets outside services (databases, browsers, CRMs) offer themselves to any compliant model as tools, with no custom glue code.
- MCP server / clientThe server offers the tools; the client (like Claude Code or claude.ai) uses them.
- Remote transport (SSE / streamable HTTP)The MCP connection for a server on a different host, with authentication. Older material says SSE; the spec has been moving toward streamable HTTP. The judgment tested: local = stdio, remote = network transport.
- Resources vs Tools (MCP)Resources are read-only things a server exposes; Tools take actions. Knowing which is which decides what an integration can and can't do.
- stdio transportThe MCP connection to use when server and client are on the same machine. Using a network transport on localhost is a classic wrong answer.
- Structured errors (isError, errorCategory, isRetryable)When a tool fails, return machine-readable error fields so the agent can plan recovery. Retry a timeout; never retry access-denied. "Something went wrong" tells the agent nothing.
- Tool descriptionThe name and description are how the model decides when to use a tool. Write them like API documentation, including a disambiguation rule for when to use THIS tool instead of a similar one.
Claude shelf
Governance & Guardrails
5 terms on this shelf.
- Approval gateA required human yes before an agent's action takes effect: sending, spending, publishing, deleting. The loop pauses until a person signs off.
- GuardrailsThe technical limits that stop an AI system from doing what policy forbids: permission rules, blocked actions, filtered inputs and outputs. Policy says what; guardrails make sure.
- HallucinationWhen a model states something false with full confidence, including invented citations, numbers, or names. It isn't lying; it's filling gaps convincingly. This is the reason verification and human review exist.
- Human in the loopA person stays close enough to an AI system's work to check it, correct it, and decide when it can run on its own versus when a person must step back in. It is a design choice about where human judgment sits in a process, not a slogan.
- Least privilegeGive an AI system the smallest access it needs and nothing more.
Claude shelf
Certification
5 terms on this shelf.
- CCAR-F / CCDV-F / CCAO-F / CCAR-PThe exam codes for Anthropic's four certifications: Architect Foundations, Developer Foundations, Associate Foundations, and Architect Professional.
- Claude Partner Network (CPN)Anthropic's partner program. Tiers (Select and up) require active certified individuals; Architect and Developer certs count, Associate does not.
- Minimally qualified candidate (MQC)The person the exam is calibrated to: someone who just barely deserves to pass. The standard-setting anchor behind the cut score.
- Pearson VUE / OnVUEThe proctored testing provider: test centers, or OnVUE for taking it from home with a webcam proctor.
- Scaled scoreExam results reported on a 100 to 1,000 scale; 720 passes. Roughly 43 to 45 of 60 questions, so about 17 wrong answers still pass.
Copilot shelf
Governance & Guardrails
5 terms on this shelf.
- AI governanceThe rules for how an organization uses AI: who may use it, for what, with which data, and who is accountable when it goes wrong. Guardrails are how those rules get enforced.
- Data sensitivityKnowing which information must never reach an AI tool (regulated personal data, client confidences, secrets) and anonymizing or withholding it before use. The redact-before-upload habit.
- GuardrailsThe technical limits that stop an AI system from doing what policy forbids: permission rules, blocked actions, filtered inputs and outputs. Policy says what; guardrails make sure.
- HallucinationWhen a model states something false with full confidence, including invented citations, numbers, or names. It isn't lying; it's filling gaps convincingly. This is the reason verification and human review exist.
- Human in the loopA person stays close enough to an AI system's work to check it, correct it, and decide when it can run on its own versus when a person must step back in. It is a design choice about where human judgment sits in a process, not a slogan.
Copilot shelf
Context & Reliability
4 terms on this shelf.
- Context windowEverything the model can see at once: the conversation, documents, tool results. Finite, and managing it is a real skill.
- LLM (Large Language Model)The engine under Claude, Copilot, and their peers: a model trained on enormous amounts of text that predicts language, which turns out to be enough to draft, summarize, reason, and converse.
- Lost in the middleModels pay the most attention to the start and end of a long context. Critical facts buried in the middle get missed.
- TokenThe unit models actually read and write: chunks of a few characters each. Context windows, pricing, and limits are all measured in tokens, roughly three quarters of a word apiece.
Copilot shelf
Agents & Orchestration
2 terms on this shelf.
- API (Application Programming Interface)The doorway a program uses to talk to another program or service over the web: send a request, get a response. Claude itself is reached through one.
- SDK (Software Development Kit)A code library that wraps an API so developers do not have to hand-build every request. Anthropic ships official Python and TypeScript SDKs for Claude.
Copilot shelf
Prompts & Structured Output
2 terms on this shelf.
Microsoft shelf
Governance & Guardrails
5 terms on this shelf.
- AI governanceThe rules for how an organization uses AI: who may use it, for what, with which data, and who is accountable when it goes wrong. Guardrails are how those rules get enforced.
- Audit trailThe record of what an AI system did and why: every action, tool call, and decision logged so a person can reconstruct it later. For compliance purposes, if it isn't logged, it didn't happen.
- Data sensitivityKnowing which information must never reach an AI tool (regulated personal data, client confidences, secrets) and anonymizing or withholding it before use. The redact-before-upload habit.
- Human in the loopA person stays close enough to an AI system's work to check it, correct it, and decide when it can run on its own versus when a person must step back in. It is a design choice about where human judgment sits in a process, not a slogan.
- Least privilegeGive an AI system the smallest access it needs and nothing more.
Microsoft shelf
Agents & Orchestration
2 terms on this shelf.
- API (Application Programming Interface)The doorway a program uses to talk to another program or service over the web: send a request, get a response. Claude itself is reached through one.
- SDK (Software Development Kit)A code library that wraps an API so developers do not have to hand-build every request. Anthropic ships official Python and TypeScript SDKs for Claude.
Microsoft shelf
Context & Reliability
2 terms on this shelf.
- LLM (Large Language Model)The engine under Claude, Copilot, and their peers: a model trained on enormous amounts of text that predicts language, which turns out to be enough to draft, summarize, reason, and converse.
- TokenThe unit models actually read and write: chunks of a few characters each. Context windows, pricing, and limits are all measured in tokens, roughly three quarters of a word apiece.
General GenAI shelf
Agents & Orchestration
12 terms on this shelf.
- Adaptive decompositionSplitting open-ended work (research, investigation) where findings reshape what the next subtask should be.
- AgentAn AI model running in a loop with tools: it thinks, takes an action, looks at the result, and goes again until the job is done.
- Agentic loopThe cycle your code runs around a stateless model: gather context, take action, verify results, repeat.
- API (Application Programming Interface)The doorway a program uses to talk to another program or service over the web: send a request, get a response. Claude itself is reached through one.
- Coordinator (orchestrator)The manager agent that breaks a big job into pieces, hands them to subagents, and combines the results.
- Hub and spokeThe orchestration shape where workers talk only to the manager, never to each other. Keeps the system predictable and debuggable.
- Refinement loopA pass where an agent's work is checked for coverage and gaps before the final version ships.
- SDK (Software Development Kit)A code library that wraps an API so developers do not have to hand-build every request. Anthropic ships official Python and TypeScript SDKs for Claude.
- Sequential decompositionSplitting a job into steps where each needs the previous step's output. Use when order matters.
- StatelessThe model keeps no memory between API calls. Every call starts blank; your code supplies the history each time.
- SubagentA worker agent spawned for one piece of a job. Returns a compact summary to the coordinator, which protects the coordinator's context budget.
- Targeted retryWhen one subagent fails, retry only that one. A blanket retry re-runs steps that already succeeded, re-firing their side effects (like duplicate database writes).
General GenAI shelf
Governance & Guardrails
10 terms on this shelf.
- AI governanceThe rules for how an organization uses AI: who may use it, for what, with which data, and who is accountable when it goes wrong. Guardrails are how those rules get enforced.
- Approval gateA required human yes before an agent's action takes effect: sending, spending, publishing, deleting. The loop pauses until a person signs off.
- Audit trailThe record of what an AI system did and why: every action, tool call, and decision logged so a person can reconstruct it later. For compliance purposes, if it isn't logged, it didn't happen.
- Data sensitivityKnowing which information must never reach an AI tool (regulated personal data, client confidences, secrets) and anonymizing or withholding it before use. The redact-before-upload habit.
- GuardrailsThe technical limits that stop an AI system from doing what policy forbids: permission rules, blocked actions, filtered inputs and outputs. Policy says what; guardrails make sure.
- HallucinationWhen a model states something false with full confidence, including invented citations, numbers, or names. It isn't lying; it's filling gaps convincingly. This is the reason verification and human review exist.
- Human in the loopA person stays close enough to an AI system's work to check it, correct it, and decide when it can run on its own versus when a person must step back in. It is a design choice about where human judgment sits in a process, not a slogan.
- Human on the loopThe lighter version of human in the loop: the AI acts on its own and a person reviews afterward, sampling results and stepping in when something looks off. In the path versus watching the path.
- Least privilegeGive an AI system the smallest access it needs and nothing more.
- Red teamingDeliberately attacking your own AI system: prompting it to misbehave, leak data, or break its rules, so you find the failure before a stranger does.
General GenAI shelf
Context & Reliability
8 terms on this shelf.
- Context budgetTreating context like money: spend it on what matters, don't dump raw tool output into it, make subagents return summaries.
- Context windowEverything the model can see at once: the conversation, documents, tool results. Finite, and managing it is a real skill.
- Escalation triggersObjective, checkable conditions for handing off to a human: the customer asked, the request is outside policy, or no progress is being made. Never the model's detected sentiment or self-reported confidence.
- Fact block / case blockA pinned, structured block of durable facts (customer, order number, issue) re-anchored at the end of the context every turn, instead of hoping the model re-finds them on page 40.
- LLM (Large Language Model)The engine under Claude, Copilot, and their peers: a model trained on enormous amounts of text that predicts language, which turns out to be enough to draft, summarize, reason, and converse.
- Lost in the middleModels pay the most attention to the start and end of a long context. Critical facts buried in the middle get missed.
- ScratchpadHaving a long-running agent write intermediate notes to a file instead of holding everything in context.
- TokenThe unit models actually read and write: chunks of a few characters each. Context windows, pricing, and limits are all measured in tokens, roughly three quarters of a word apiece.
General GenAI shelf
Prompts & Structured Output
5 terms on this shelf.
- Confidence calibration (field-level)In extraction pipelines, scoring confidence per field (99% on the date, 40% on the vendor) so low-confidence fields route to human review. Different from escalating a conversation on the model's mood, which is always wrong.
- Few-shot examplesShowing 2 to 4 examples of the output you want inside the prompt. Beats describing the format in words.
- Mechanically checkable rulesReplacing vague adjectives ("be careful") with categorical rules a checker could verify. Vague adjectives produce false positives.
- Nullable fieldsMarking schema fields optional so "not present" is a legal answer. A required field with missing source data gets fabricated. Required + missing = made up.
- Validation-retry loopCheck structured output against the schema; if invalid, send the error back for a corrected retry (usually succeeds), then fail loudly. Never silently accept bad data.
General GenAI shelf
Tools & MCP
3 terms on this shelf.
- MCP (Model Context Protocol)The open standard that lets outside services (databases, browsers, CRMs) offer themselves to any compliant model as tools, with no custom glue code.
- MCP server / clientThe server offers the tools; the client (like Claude Code or claude.ai) uses them.
- Tool descriptionThe name and description are how the model decides when to use a tool. Write them like API documentation, including a disambiguation rule for when to use THIS tool instead of a similar one.
General GenAI shelf
Certification
3 terms on this shelf.
- Minimally qualified candidate (MQC)The person the exam is calibrated to: someone who just barely deserves to pass. The standard-setting anchor behind the cut score.
- Pearson VUE / OnVUEThe proctored testing provider: test centers, or OnVUE for taking it from home with a webcam proctor.
- Scaled scoreExam results reported on a 100 to 1,000 scale; 720 passes. Roughly 43 to 45 of 60 questions, so about 17 wrong answers still pass.
General GenAI shelf
Claude Code
1 term on this shelf.
Human in the Loop
The namesake shelf: governance, guardrails, and keeping people in charge of AI.
AI governanceThe rules for how an organization uses AI: who may use it, for what, with which data, and who is accountable when it goes wrong. Guardrails are how those rules get enforced.Approval gateA required human yes before an agent's action takes effect: sending, spending, publishing, deleting. The loop pauses until a person signs off.Audit trailThe record of what an AI system did and why: every action, tool call, and decision logged so a person can reconstruct it later. For compliance purposes, if it isn't logged, it didn't happen.Confidence calibration (field-level)In extraction pipelines, scoring confidence per field (99% on the date, 40% on the vendor) so low-confidence fields route to human review. Different from escalating a conversation on the model's mood, which is always wrong.dangerously-skip-permissionsThe flag that bypasses every permission prompt. Never use it, except in strictly controlled, low-risk environments like a locked-down CI job.Data sensitivityKnowing which information must never reach an AI tool (regulated personal data, client confidences, secrets) and anonymizing or withholding it before use. The redact-before-upload habit.Escalation triggersObjective, checkable conditions for handing off to a human: the customer asked, the request is outside policy, or no progress is being made. Never the model's detected sentiment or self-reported confidence.GuardrailsThe technical limits that stop an AI system from doing what policy forbids: permission rules, blocked actions, filtered inputs and outputs. Policy says what; guardrails make sure.HallucinationWhen a model states something false with full confidence, including invented citations, numbers, or names. It isn't lying; it's filling gaps convincingly. This is the reason verification and human review exist.Human in the loopA person stays close enough to an AI system's work to check it, correct it, and decide when it can run on its own versus when a person must step back in. It is a design choice about where human judgment sits in a process, not a slogan.Human on the loopThe lighter version of human in the loop: the AI acts on its own and a person reviews afterward, sampling results and stepping in when something looks off. In the path versus watching the path.Least privilegeGive an AI system the smallest access it needs and nothing more.Permission rules (allow / ask / deny)Per-tool access control for Claude Code: allow silently, ask first, or deny outright, for tools like bash, read, edit, and web fetch.Red teamingDeliberately attacking your own AI system: prompting it to misbehave, leak data, or break its rules, so you find the failure before a stranger does.SandboxingRestricting what an agent's shell can touch on the host machine (tools like bubblewrap and socat). Containment beats trust.
Claude
Anthropic's Claude: agents, Claude Code, and the architecture behind them.
.claude folderThe project's Claude configuration home: rules/ (always-on instructions), commands/ (reusable /name prompts), skills/ (bigger packaged abilities)..mcp.json vs ~/.claude.jsonWhere tool connections live: .mcp.json in the project is shared with the team; ~/.claude.json is personal. Same logic as project vs personal CLAUDE.md.@claude (GitHub Actions)Mentioning @claude on a GitHub issue or PR to trigger automated review or fixes through a GitHub Action.@importHow a CLAUDE.md pulls other files into itself, so instructions can be split up but load together.Adaptive decompositionSplitting open-ended work (research, investigation) where findings reshape what the next subtask should be.AgentAn AI model running in a loop with tools: it thinks, takes an action, looks at the result, and goes again until the job is done.Agent SDKAnthropic's framework that runs the agent loop for you: manages subagents, ships built-in tools. The raw API makes you write the loop yourself.Agentic loopThe cycle your code runs around a stateless model: gather context, take action, verify results, repeat.API (Application Programming Interface)The doorway a program uses to talk to another program or service over the web: send a request, get a response. Claude itself is reached through one.Approval gateA required human yes before an agent's action takes effect: sending, spending, publishing, deleting. The loop pauses until a person signs off.Call-inspect-execute-append cycleThe full agent turn: call the API, inspect the stop_reason, execute the requested tool locally, append the result to the history, call again. Skip inspect and the agent goes silent; skip append and it repeats the same tool call forever.CCAR-F / CCDV-F / CCAO-F / CCAR-PThe exam codes for Anthropic's four certifications: Architect Foundations, Developer Foundations, Associate Foundations, and Architect Professional.Claude CodeAnthropic's terminal agent. It works inside a folder or repository with real file access through tools like bash and grep.Claude Partner Network (CPN)Anthropic's partner program. Tiers (Select and up) require active certified individuals; Architect and Developer certs count, Associate does not.CLAUDE.mdThe instruction file Claude Code reads at session start. Three levels: user (~/.claude/, personal, never version-controlled), project (repo root, committed, team rules), and directory (subfolder-specific rules).Confidence calibration (field-level)In extraction pipelines, scoring confidence per field (99% on the date, 40% on the vendor) so low-confidence fields route to human review. Different from escalating a conversation on the model's mood, which is always wrong.Context budgetTreating context like money: spend it on what matters, don't dump raw tool output into it, make subagents return summaries.Context windowEverything the model can see at once: the conversation, documents, tool results. Finite, and managing it is a real skill.Coordinator (orchestrator)The manager agent that breaks a big job into pieces, hands them to subagents, and combines the results.dangerously-skip-permissionsThe flag that bypasses every permission prompt. Never use it, except in strictly controlled, low-risk environments like a locked-down CI job.end_turnA stop_reason value meaning "I'm done." The loop ends here, and only here.Escalation triggersObjective, checkable conditions for handing off to a human: the customer asked, the request is outside policy, or no progress is being made. Never the model's detected sentiment or self-reported confidence.Fact block / case blockA pinned, structured block of durable facts (customer, order number, issue) re-anchored at the end of the context every turn, instead of hoping the model re-finds them on page 40.Few-shot examplesShowing 2 to 4 examples of the output you want inside the prompt. Beats describing the format in words.fork_sessionStarts a new session as a copy of an existing session's context, so parallel work can branch from the same starting point.GuardrailsThe technical limits that stop an AI system from doing what policy forbids: permission rules, blocked actions, filtered inputs and outputs. Policy says what; guardrails make sure.HallucinationWhen a model states something false with full confidence, including invented citations, numbers, or names. It isn't lying; it's filling gaps convincingly. This is the reason verification and human review exist.Headless mode (-p)Runs Claude Code non-interactively with one prompt, for scripts and CI/CD. Pair with --output-format json for machine-readable results.HookA small script that runs automatically at a fixed moment in an agent's work. Rules that must ALWAYS happen belong in hooks. Prompts are requests; hooks are law.Hub and spokeThe orchestration shape where workers talk only to the manager, never to each other. Keeps the system predictable and debuggable.Human in the loopA person stays close enough to an AI system's work to check it, correct it, and decide when it can run on its own versus when a person must step back in. It is a design choice about where human judgment sits in a process, not a slogan.JSON schema (as a tool)Defining the exact shape of data you need as a tool definition, so the model fills in a form instead of writing an essay you have to parse.Least privilegeGive an AI system the smallest access it needs and nothing more.LLM (Large Language Model)The engine under Claude, Copilot, and their peers: a model trained on enormous amounts of text that predicts language, which turns out to be enough to draft, summarize, reason, and converse.Lost in the middleModels pay the most attention to the start and end of a long context. Critical facts buried in the middle get missed.MCP (Model Context Protocol)The open standard that lets outside services (databases, browsers, CRMs) offer themselves to any compliant model as tools, with no custom glue code.MCP server / clientThe server offers the tools; the client (like Claude Code or claude.ai) uses them.Mechanically checkable rulesReplacing vague adjectives ("be careful") with categorical rules a checker could verify. Vague adjectives produce false positives.Message Batches APISubmit thousands of independent prompts as an offline batch: results within 24 hours at 50% of the cost. Never for anything a person is waiting on.Minimally qualified candidate (MQC)The person the exam is calibrated to: someone who just barely deserves to pass. The standard-setting anchor behind the cut score.Nullable fieldsMarking schema fields optional so "not present" is a legal answer. A required field with missing source data gets fabricated. Required + missing = made up.Pearson VUE / OnVUEThe proctored testing provider: test centers, or OnVUE for taking it from home with a webcam proctor.Permission rules (allow / ask / deny)Per-tool access control for Claude Code: allow silently, ask first, or deny outright, for tools like bash, read, edit, and web fetch.Plan modeMakes Claude Code propose an approach for approval before touching anything. Right for big risky changes; overkill for one-liners.PreToolUse / PostToolUseThe two classic hook moments: right before a tool runs (gate it) and right after (enforce follow-ups, like logging every change).Prompt cachingMarking static sections (system prompt, few-shot examples) with cache breakpoints so repeat calls reuse them, cutting cost up to ~90% on the cached portion.Refinement loopA pass where an agent's work is checked for coverage and gaps before the final version ships.Remote transport (SSE / streamable HTTP)The MCP connection for a server on a different host, with authentication. Older material says SSE; the spec has been moving toward streamable HTTP. The judgment tested: local = stdio, remote = network transport.Resources vs Tools (MCP)Resources are read-only things a server exposes; Tools take actions. Knowing which is which decides what an integration can and can't do.SandboxingRestricting what an agent's shell can touch on the host machine (tools like bubblewrap and socat). Containment beats trust.Scaled scoreExam results reported on a 100 to 1,000 scale; 720 passes. Roughly 43 to 45 of 60 questions, so about 17 wrong answers still pass.ScratchpadHaving a long-running agent write intermediate notes to a file instead of holding everything in context.SDK (Software Development Kit)A code library that wraps an API so developers do not have to hand-build every request. Anthropic ships official Python and TypeScript SDKs for Claude.Sequential decompositionSplitting a job into steps where each needs the previous step's output. Use when order matters.Session commands (resume, fork, compact, clear, rewind)Managing a session's life: re-enter an old one (resume), branch its history (fork), shrink token use (compact), wipe it (clear), roll back to an earlier turn (rewind).SKILL.md frontmatterThe header block of a skill file. Can set context: fork, allowed-tools, and argument-hint.Slash commandA reusable prompt saved as a markdown file in .claude/commands, triggered by typing /name. Its YAML frontmatter can restrict tools and hint at arguments.StatelessThe model keeps no memory between API calls. Every call starts blank; your code supplies the history each time.stdio transportThe MCP connection to use when server and client are on the same machine. Using a network transport on localhost is a classic wrong answer.stop_reasonThe machine-readable label on every API reply saying why the model stopped. Your loop should be driven by this label, never by reading the model's prose or counting turns.Structured errors (isError, errorCategory, isRetryable)When a tool fails, return machine-readable error fields so the agent can plan recovery. Retry a timeout; never retry access-denied. "Something went wrong" tells the agent nothing.SubagentA worker agent spawned for one piece of a job. Returns a compact summary to the coordinator, which protects the coordinator's context budget.Targeted retryWhen one subagent fails, retry only that one. A blanket retry re-runs steps that already succeeded, re-firing their side effects (like duplicate database writes).Task toolHow a coordinator in the Agent SDK/Claude Code actually spawns a subagent.The four surfacesThe map of the Claude ecosystem: the Claude API (one stateless HTTP endpoint plus SDKs), the Agent SDK (runs the loop), Claude Code (a terminal agent with file access), and MCP (the standard plug for outside tools).TokenThe unit models actually read and write: chunks of a few characters each. Context windows, pricing, and limits are all measured in tokens, roughly three quarters of a word apiece.Tool descriptionThe name and description are how the model decides when to use a tool. Write them like API documentation, including a disambiguation rule for when to use THIS tool instead of a similar one.tool_choicePer-request control of tool use: auto (model decides), any (must use some tool), a named tool (must use that one), or none.tool_useA stop_reason value meaning "run this tool for me and send back the result." The loop continues.Validation-retry loopCheck structured output against the schema; if invalid, send the error back for a corrected retry (usually succeeds), then fail loudly. Never silently accept bad data.
Copilot
Microsoft Copilot terminology. This shelf is growing.
AI governanceThe rules for how an organization uses AI: who may use it, for what, with which data, and who is accountable when it goes wrong. Guardrails are how those rules get enforced.API (Application Programming Interface)The doorway a program uses to talk to another program or service over the web: send a request, get a response. Claude itself is reached through one.Context windowEverything the model can see at once: the conversation, documents, tool results. Finite, and managing it is a real skill.Data sensitivityKnowing which information must never reach an AI tool (regulated personal data, client confidences, secrets) and anonymizing or withholding it before use. The redact-before-upload habit.Few-shot examplesShowing 2 to 4 examples of the output you want inside the prompt. Beats describing the format in words.GuardrailsThe technical limits that stop an AI system from doing what policy forbids: permission rules, blocked actions, filtered inputs and outputs. Policy says what; guardrails make sure.HallucinationWhen a model states something false with full confidence, including invented citations, numbers, or names. It isn't lying; it's filling gaps convincingly. This is the reason verification and human review exist.Human in the loopA person stays close enough to an AI system's work to check it, correct it, and decide when it can run on its own versus when a person must step back in. It is a design choice about where human judgment sits in a process, not a slogan.LLM (Large Language Model)The engine under Claude, Copilot, and their peers: a model trained on enormous amounts of text that predicts language, which turns out to be enough to draft, summarize, reason, and converse.Lost in the middleModels pay the most attention to the start and end of a long context. Critical facts buried in the middle get missed.Mechanically checkable rulesReplacing vague adjectives ("be careful") with categorical rules a checker could verify. Vague adjectives produce false positives.SDK (Software Development Kit)A code library that wraps an API so developers do not have to hand-build every request. Anthropic ships official Python and TypeScript SDKs for Claude.TokenThe unit models actually read and write: chunks of a few characters each. Context windows, pricing, and limits are all measured in tokens, roughly three quarters of a word apiece.
Microsoft
The wider Microsoft AI stack. Books arriving.
AI governanceThe rules for how an organization uses AI: who may use it, for what, with which data, and who is accountable when it goes wrong. Guardrails are how those rules get enforced.API (Application Programming Interface)The doorway a program uses to talk to another program or service over the web: send a request, get a response. Claude itself is reached through one.Audit trailThe record of what an AI system did and why: every action, tool call, and decision logged so a person can reconstruct it later. For compliance purposes, if it isn't logged, it didn't happen.Data sensitivityKnowing which information must never reach an AI tool (regulated personal data, client confidences, secrets) and anonymizing or withholding it before use. The redact-before-upload habit.Human in the loopA person stays close enough to an AI system's work to check it, correct it, and decide when it can run on its own versus when a person must step back in. It is a design choice about where human judgment sits in a process, not a slogan.Least privilegeGive an AI system the smallest access it needs and nothing more.LLM (Large Language Model)The engine under Claude, Copilot, and their peers: a model trained on enormous amounts of text that predicts language, which turns out to be enough to draft, summarize, reason, and converse.SDK (Software Development Kit)A code library that wraps an API so developers do not have to hand-build every request. Anthropic ships official Python and TypeScript SDKs for Claude.TokenThe unit models actually read and write: chunks of a few characters each. Context windows, pricing, and limits are all measured in tokens, roughly three quarters of a word apiece.
General GenAI
Concepts that apply no matter whose AI you use.
Adaptive decompositionSplitting open-ended work (research, investigation) where findings reshape what the next subtask should be.AgentAn AI model running in a loop with tools: it thinks, takes an action, looks at the result, and goes again until the job is done.Agentic loopThe cycle your code runs around a stateless model: gather context, take action, verify results, repeat.AI governanceThe rules for how an organization uses AI: who may use it, for what, with which data, and who is accountable when it goes wrong. Guardrails are how those rules get enforced.API (Application Programming Interface)The doorway a program uses to talk to another program or service over the web: send a request, get a response. Claude itself is reached through one.Approval gateA required human yes before an agent's action takes effect: sending, spending, publishing, deleting. The loop pauses until a person signs off.Audit trailThe record of what an AI system did and why: every action, tool call, and decision logged so a person can reconstruct it later. For compliance purposes, if it isn't logged, it didn't happen.Confidence calibration (field-level)In extraction pipelines, scoring confidence per field (99% on the date, 40% on the vendor) so low-confidence fields route to human review. Different from escalating a conversation on the model's mood, which is always wrong.Context budgetTreating context like money: spend it on what matters, don't dump raw tool output into it, make subagents return summaries.Context windowEverything the model can see at once: the conversation, documents, tool results. Finite, and managing it is a real skill.Coordinator (orchestrator)The manager agent that breaks a big job into pieces, hands them to subagents, and combines the results.Data sensitivityKnowing which information must never reach an AI tool (regulated personal data, client confidences, secrets) and anonymizing or withholding it before use. The redact-before-upload habit.Escalation triggersObjective, checkable conditions for handing off to a human: the customer asked, the request is outside policy, or no progress is being made. Never the model's detected sentiment or self-reported confidence.Fact block / case blockA pinned, structured block of durable facts (customer, order number, issue) re-anchored at the end of the context every turn, instead of hoping the model re-finds them on page 40.Few-shot examplesShowing 2 to 4 examples of the output you want inside the prompt. Beats describing the format in words.GuardrailsThe technical limits that stop an AI system from doing what policy forbids: permission rules, blocked actions, filtered inputs and outputs. Policy says what; guardrails make sure.HallucinationWhen a model states something false with full confidence, including invented citations, numbers, or names. It isn't lying; it's filling gaps convincingly. This is the reason verification and human review exist.Hub and spokeThe orchestration shape where workers talk only to the manager, never to each other. Keeps the system predictable and debuggable.Human in the loopA person stays close enough to an AI system's work to check it, correct it, and decide when it can run on its own versus when a person must step back in. It is a design choice about where human judgment sits in a process, not a slogan.Human on the loopThe lighter version of human in the loop: the AI acts on its own and a person reviews afterward, sampling results and stepping in when something looks off. In the path versus watching the path.Least privilegeGive an AI system the smallest access it needs and nothing more.LLM (Large Language Model)The engine under Claude, Copilot, and their peers: a model trained on enormous amounts of text that predicts language, which turns out to be enough to draft, summarize, reason, and converse.Lost in the middleModels pay the most attention to the start and end of a long context. Critical facts buried in the middle get missed.MCP (Model Context Protocol)The open standard that lets outside services (databases, browsers, CRMs) offer themselves to any compliant model as tools, with no custom glue code.MCP server / clientThe server offers the tools; the client (like Claude Code or claude.ai) uses them.Mechanically checkable rulesReplacing vague adjectives ("be careful") with categorical rules a checker could verify. Vague adjectives produce false positives.Minimally qualified candidate (MQC)The person the exam is calibrated to: someone who just barely deserves to pass. The standard-setting anchor behind the cut score.Nullable fieldsMarking schema fields optional so "not present" is a legal answer. A required field with missing source data gets fabricated. Required + missing = made up.Pearson VUE / OnVUEThe proctored testing provider: test centers, or OnVUE for taking it from home with a webcam proctor.Red teamingDeliberately attacking your own AI system: prompting it to misbehave, leak data, or break its rules, so you find the failure before a stranger does.Refinement loopA pass where an agent's work is checked for coverage and gaps before the final version ships.SandboxingRestricting what an agent's shell can touch on the host machine (tools like bubblewrap and socat). Containment beats trust.Scaled scoreExam results reported on a 100 to 1,000 scale; 720 passes. Roughly 43 to 45 of 60 questions, so about 17 wrong answers still pass.ScratchpadHaving a long-running agent write intermediate notes to a file instead of holding everything in context.SDK (Software Development Kit)A code library that wraps an API so developers do not have to hand-build every request. Anthropic ships official Python and TypeScript SDKs for Claude.Sequential decompositionSplitting a job into steps where each needs the previous step's output. Use when order matters.StatelessThe model keeps no memory between API calls. Every call starts blank; your code supplies the history each time.SubagentA worker agent spawned for one piece of a job. Returns a compact summary to the coordinator, which protects the coordinator's context budget.Targeted retryWhen one subagent fails, retry only that one. A blanket retry re-runs steps that already succeeded, re-firing their side effects (like duplicate database writes).TokenThe unit models actually read and write: chunks of a few characters each. Context windows, pricing, and limits are all measured in tokens, roughly three quarters of a word apiece.Tool descriptionThe name and description are how the model decides when to use a tool. Write them like API documentation, including a disambiguation rule for when to use THIS tool instead of a similar one.Validation-retry loopCheck structured output against the schema; if invalid, send the error back for a corrected retry (usually succeeds), then fail loudly. Never silently accept bad data.