Approving agent actions safely

An agent that can act is only as trustworthy as the process around it. The approval queue is where staged WRITE and RISK proposals wait for a human decision, and where you build the evidence to trust an agent with more.

What a proposal contains

• The tool being called, for example reply_to_ticket or advance_deal_stage.
• The risk level: WRITE or RISK.
• The exact arguments the agent wants to use.
• A one-line summary written for the reviewer.

Review workflow

1. 1Open Cortex > Inbox (/cortex/inbox). The badge in the Agent Hub header counts actions waiting for you; Needs attention also lists approvals alongside hand-offs and failed runs.
2. 2Open a run to see its drawer: the input, each step, the tools used, credits spent and the result.
3. 3Read each proposal's summary, then check the arguments, such as the reply text or amount.
4. 4Check the cited sources and the confidence score to see whether the answer was grounded.
5. 5Approve, edit or reject each proposed action individually, or bulk-approve low-risk ones. Add feedback so the playbook can be improved. The run closes once every proposal is decided.
6. 6If several proposals in a row need the same correction, fix the playbook steps rather than repeating the edit.

Tuning knobs

• confidenceToSend: below this score the agent stages a draft instead of sending.
• spendCapCredits: the maximum credits a run may use.
• escalateRules: phrases or conditions (for example legal threats or refund requests above a limit) that force hand-off.
• toolConfig: per-tool limits such as capUsd for refunds.

Graduating autonomy

Move from DRAFT to AUTO_LOW only after a sustained run of proposals approved without edits. Keep RISK tools staged even at higher autonomy where your policy requires it.