August 5, 2026 · By Ada
Agent Instructions: What They Are, and Why an Agent Does Nothing Without Them
Agent instructions are the standing text that tells an AI agent what its job is: what to do, to what standard, and what to bring back to a human instead of deciding on its own. The agent reads them on every run. Without them it has access to everything and no reason to touch any of it.
This is the I in ITT, the three parts something needs before it is an agent. The other two are tools and triggers.
Imagine hiring someone and never writing the job description
You just hired someone. You give them the laptop, the phone, the logins, access to every system they will need. You tell them exactly what time to start on Monday.
But you never gave them a job description.
They show up. They sit down. They do nothing all day.
That is not a bad hire. That is a person who was told nothing, doing nothing, correctly, for eight hours. Nobody would blame the hire. They would go and find the job description.
That is exactly what happens to an agent with no instructions, and it is the first thing that happened to Ava Idris, our AI Chief of Staff. She had the access. She knew when to start. Nothing told her what the job was.
The three parts, and where instructions sit
An agent needs three things, and they answer three different questions.
Triggers decide when the agent starts. A time, a date, an email landing, a form somebody fills in. This is the part that runs without anyone asking. More on triggers.
Tools decide what it can reach. Integrations with the apps you already use: Gmail, Slack, your calendar, Canva. Without them an agent can describe work but cannot do any of it. More on tools.
Instructions decide what it actually does. Check the inbox and flag anything urgent. Draft the reply. Design the graphic in the brand template. This is the plain description of the job, written the way you would write it for a person.
Ava's first day had two out of three. The trigger was there, she started at nine. The tools were there, the whole calendar and inbox. The instructions were missing, so the other two had nothing to serve.
Instructions are not only for agents
This is where the word gets confusing, because instructions show up in four places and only one of them is an agent.
In a prompt. You type what you want, it runs once, it is gone. The instruction and the request are the same act.
In a project. Claude Projects, custom GPTs, a workspace with standing context. Instructions written once that apply to every chat inside it. Genuinely useful, and a real thing to set up well.
In an agent. The same standing text, except now something else starts the run and the agent has tools to act with. Nobody has to be in the chat for it to happen.
In a skill. A named procedure the agent loads when a job calls for it. Instructions for one task rather than for the whole role, so the agent can hold a lot of them without carrying all of them at once.
Here is the part worth being precise about. Having instructions does not make something an agent. A project with beautifully written instructions still sits there until you open it and type. What makes it an agent is the trigger that starts it and the tools it can act with. Instructions are necessary. They are not sufficient.
What a good instruction actually says
Not a personality. What the job is, in the words you would use with somebody in their first week.
- The job, in verbs. "Check the inbox each morning and flag anything that needs an answer today." Not "be helpful with email."
- What done looks like. Specific enough that two people would agree on whether it was met.
- What it must never do. The lines it brings to a human instead of deciding. Ours: anything that reaches more than fifty people gets read by a human first.
- The context it cannot infer. Names, preferences, and the reason something is done the odd way it is done.
Why this is the piece to write first
Tools without instructions is an agent that can act and does not know what good means. Triggers without instructions is that same agent, running on its own, on a schedule.
Instructions cost nothing but thinking time, and they are the only one of the three that makes everything downstream better. Write them before you hand anything access.
So what did you actually tell yours it was for?
What are AI agent instructions?
The standing text that tells an agent what its job is: what to do, to what standard, and what to bring to a human instead of deciding alone. The agent reads them on every run, so nobody has to restate the job each time.
Why does my AI agent do nothing?
Usually because it has access and a schedule but no description of the job. An agent told nothing does nothing, correctly. Check the instructions before assuming the agent is broken.
Do instructions make something an AI agent?
No. Instructions are one of three parts. A project with standing instructions still waits for you to open it and type. It becomes an agent when it also has tools it can act with and a trigger that starts it without you.
Where do agent instructions live?
In four common places: a prompt, a project's standing context, an agent's own configuration, or a skill the agent loads for a specific task. What matters is that the agent reads them on every run.
What should agent instructions include?
The job written in verbs, what finished work looks like, what the agent must never do or must escalate to a human, and the context it cannot infer on its own such as names, preferences and why something is done an unusual way.