Configuring your agents

Each employee on your AI Workforce is tunable per product, without touching a prompt. Open a product → AI Workforce, pick a colleague, then use the Guidelines tab and the gear drawer.

Guidelines — how it should behave

The Guidelines tab asks each role the questions it actually needs answered, grouped by topic ("Tone of voice", "Escalation", …). Your one-line answers ride along on every prompt that employee sends, so behaviour is tuned in plain English rather than by rewriting instructions.

Guidelines apply from the next message onward. A rule saved under a group that employee doesn't have is dropped — check it landed under one of its own groups if it seems ignored.

The gear drawer — what it is and what it may do

  • Name and persona — rename an employee and give it standing instructions.
  • Abilities — a switch per tool. Turning one off genuinely removes it from that employee's toolset on the next turn; it is not a hint.
  • Briefings — off by default. Set a daily or weekly cadence and that employee reports on its own into a dedicated Briefings conversation.
  • Active — retire an employee you don't want in the team rail.

There is no per-employee model picker, and that is deliberate: a model id is valid for exactly one lab, so per-agent model fields are how a deployment ends up pointing one provider's id at another provider and failing every turn. Which model answers is an operator decision, expressed as routing: every call names the task it performs, and the task maps to a fast / balanced / deep tier that the workspace operator configures once.

Testing before you trust it

Every write an agent proposes is confirmation-gated — it records a pending action you approve or discard, and nothing mutates until you approve. That is the test harness: let it propose, read the exact payload, then decide.

The Activity tab is the audit trail — every action ever proposed, what happened to it, and everything the team has remembered about this product. Delete any memory that is wrong; memories are injected into future conversations.

Why it matters

Approval-gating plus a per-employee ability switch means an agent's blast radius is something you set, not something you hope for.