First agent guide
From a described job to a deployed one.
The first agent's purpose is not to save the most time. It is to teach you what a run timeline looks like when something is wrong, cheaply — so pick accordingly.
Pick something repetitive and low consequence
A morning briefing that reads your systems and emails four managers is close to ideal: real value, no consequential action, and it exercises reads across several connected systems. The instinct is to start with the most expensive process, which means spending the first fortnight nervous instead of learning.
Write the job in one sentence
If it does not fit in a sentence with a schedule, a condition and an escalation, it is probably two agents. "Every weekday at seven, summarise what is stuck and what needs a decision, and send it to the management team" is a complete specification.
Read the review screen properly, once
It is the only screen on which you accept the risk, and the only time you will read it carefully rather than skimming. Check the read list against the job, the action list against what you expect, and the boundary list for anything that surprises you.
Dry run against last week before you deploy
"It would have emailed 17 customers and asked for 4 approvals" is worth more than any amount of description, and it occasionally reveals that your data is not what you assumed. A dry run against a quiet week proves little — pick a busy one.
Watch the first fortnight, then stop watching
Read every run for two weeks. You are looking for three things: decisions you disagree with, escalations that should have been handled, and things handled that should have escalated. After that, switch to reading only the exceptions — continuing to watch everything is how an agent becomes a second job.
- Decisions you would have made differently
- Escalations the agent could safely have handled
- Actions it took that should have been escalated
- Anything it read that you did not expect it to need
Then widen it, one capability at a time
Add capabilities deliberately, one per change, and give each a fortnight. An agent that gained four capabilities in a week cannot be evaluated, because you no longer know which change caused the behaviour you are looking at.
Give AI a job.
Build an agent, choose what it can access, define what it may do, and deploy it to work safely across your business.
No code required · Full SDKs and API · Stop any agent instantly · [email protected]