Working with agents

The day-to-day loop with an agent is conversational: you ask in chat, it works, and you can watch every step it took, live or after the fact.

Mention an agent

@-mention an agent in any channel or conversation and it joins with full context: the discussion so far, the attached issues and PRs, the repository in scope. Asking the triage agent to "have a look at this" in an issue's conversation gives it the issue, not a vacuum.

DM an agent

A direct message to an agent is a private working session. Use it for tasks that do not belong in a channel: "summarise what changed in the payment flow this month", "draft an issue for the flakiness we discussed". The agent replies in the DM and its actions land on the platform, attributed.

Sessions and replay

Every agent run is a session. The Agents view lists them all: which agent, who started it, status (running, ended, failed, revoked), and cost. Open a session for the replay: each step the agent took, each tool it called, each write it made, in order. Live sessions stream; finished ones are a permanent record.

This is the difference between automation you trust and automation you tolerate: when an agent does something surprising, you read exactly what happened instead of guessing.

Pausing and revoking

A running session can be revoked from its page, which stops it immediately. Sessions also end themselves: when the task completes, when they hit an error, or when they reach the organisation's spend limits.