Agents in the console
What an agent is, its YAML and its runtime: Agents, Agent definition and Agent runtime. This page covers the screens.
Agent directory (#/agents)
One row per agent. Click a row, or move with ↑ / ↓ and press Enter or Space, to open its page. New agent creates one.
| Column | Meaning |
|---|---|
| Status | Enabled (green, accepts sessions), Disabled (shown as paused: no new sessions) or Needs setup (the definition has an issue, such as an MCP server that is not connected, and cannot be enabled until fixed). A line under the table explains what each means |
| Reach | How people get to the agent: the web chat link (open or copy), the channel provider with its connection (email, iMessage, Slack, Telegram), the schedule, or "Console only". Search matches reach too |
| Tools / Approvals | The first few tools and ask-first approvals. View all opens a dialog with the rest |
| Sessions, cost | Sessions in the last 7 days; all-time cost per agent. The page header shows total sessions and total cost across all agents (a dash when no cost is known) |
Filter chips: All, Enabled, Disabled, Needs setup.
Create (#/agents/new)
A form beside an author chat. Describe the agent; the author fills the form and you review it. The page shows "Ready to save" when name and instructions are present. Draft YAML previews the file. A new agent is always created disabled.
Agent page (#/agents/<id>)
Header buttons: Edit agent, and Enable agent or Disable agent.
- Disable agent asks first. The agent stops accepting new work: no new sessions, and its web chat page and channel stop answering. A turn already running finishes; a waiting approval can only be denied. Nothing is deleted, and you can enable it again.
- A web chat card shows the link state: Live (share the link), Waiting for you to enable the agent, or Off (turn it on in Edit agent). See Web chat.
- A banner appears when the YAML file was edited into an invalid state; the last valid revision stays in use.
Tabs:
| Tab | Shows |
|---|---|
| Sessions | This agent's sessions, with filters (status: Open, Running, Waiting on you, Held, ...; source) and stats: active runs, sessions this week, run success rate, estimated cost |
| Configuration | The current definition, read-only |
Edit (#/agents/<id>/edit)
The definition form with a tool checklist (builtin groups, MCP servers and their tools, skills, Stageflow tools), approvals, limits, channel and schedule, web chat and observability. Saving shows a diff review against the current revision and version history (every save is a revision; restore an older one). ⌘S saves. A stale edit is refused as agent.conflict. Remove archives the agent (agents/<id>.agent.yaml leaves the project; history stays readable).
Sessions
- Agent sessions (
#/agent-sessions) lists every agent's sessions with search and filters; the rail badge counts sessions in flight.#/agents/sessionsopens the same page. - A session (
#/agents/<id>/sessions/<sessionId>) is the conversation, with a composer to message the agent as an operator, the turns, any pending approval (Approve / Deny), delivery state for channel replies, and Stop.
Run inspector (#/agents/<id>/sessions/<sessionId>/runs/<runId>)
One turn of a session in five tabs:
| Tab | Shows |
|---|---|
| Steps | The run as a readable story, with Approve and Deny on a pending approval |
| Waterfall | Spans on a timeline, with the span filter and a detail panel |
| Messages | The messages in the turn |
| Logs | Built from spans, approvals and errors (no log records are stored) |
| Configuration | The revision the run used |
Spans and capture modes: Agent observability.
Approvals
A pending approval appears in three places: the Inbox (and the desktop notification, see Settings), the session page and the run inspector. The same decision is also available through the REST API, CLI and MCP.