An agent is a digital colleague for one recurring job: reading certificate documents as they arrive, collecting the evidence a delivery needs, or chasing a supplier for a renewed certificate. It reads, matches and drafts. It never changes your records by itself: what it prepares waits for a person.
The playbook
Each agent follows a playbook drawn on a canvas in Agents, from top to bottom:
- When: what starts a run. Someone chooses Run an agent on a lot or a certificate document, or something happens that an automation watches, such as a lot delivered or a certificate about to expire. Only if limits it to events that meet conditions, such as a receiver in the EU.
- Step: the model follows one instruction, using only the tools you allow it, and returns the fields you name. Steps only read and prepare.
- Rule or Judgement: where the path splits. A rule checks fields the run already knows; a judgement asks the fast model one short question.
- Person reviews: a person checks a draft, an answer or fields read from a document before the run goes on. Every message is reviewed before it leaves.
- Approval: every proposal waits in Approvals for a person with approval rights.
- Result: what the run records at the end.
Write the agent's Goal and your company's own rules in the panel on the right, as you would tell a new colleague. Every step reads them.
Test before anything runs
Add at least one scenario: what starts the run, what each tool answers and the path the run must take. Test runs every scenario, and six safety cases in each of them, with made-up data. Nothing is sent or proposed. A safety case passes only when the run takes the same path as its scenario. After a platform update changes the safety cases, the version in use is tested again; if it fails, it pauses. The canvas then shows Last test: how many runs took each path.
When a case fails, Last test explains it in plain words: what we tried, what the agent did wrong, why it matters and what to change, naming the step by its title. Show on the playbook draws the route the run took (solid) and the route it should have taken (dashed). Show me points at the step to change; when the change is always the same, Fix it makes it on the draft. Save the draft and test again. When a draft comes from an older version of a template, a notice lists what the newer version changes; Update this draft replaces its playbook and scenarios and keeps its name, goal and rules.
Trial, then in use
An owner or admin puts a tested draft on trial with Start trial. On trial, the agent works on its real triggers, but nothing is sent and nothing is proposed. What it would have done waits in Review for people to mark Right or Wrong. After at least 7 days and 20 marked runs, with enough of them right and every person's step reached, an owner or admin chooses Put in use.
Keep it safe
- Each run has limits: nodes, tool calls, proposals, time and cost. A monthly cost limit pauses the agent when it is reached.
- An agent that people reject too often, or that reads suspicious text too often, pauses by itself.
- Pause stops one agent. In Settings → AI & data, Halt all agents stops every agent of the company at once.
An agent never approves, never changes settings, members or keys, and never acts for a person. It only proposes.