Skip to content

Latest commit

 

History

History
83 lines (55 loc) · 2.76 KB

agents.md

File metadata and controls

83 lines (55 loc) · 2.76 KB

Agents

Baudbot uses a small multi-agent architecture: one persistent orchestrator and task-scoped workers. This keeps the team-facing interface stable while allowing parallel execution in the background.

Role overview

Agent Lifecycle Primary responsibility
control-agent Persistent Intake, triage, delegation, user communication, orchestration
dev-agent Ephemeral per task Code changes, tests, PRs, CI/review iteration
sentry-agent Persistent/on-demand Sentry alert triage and investigation support

Control-agent

The control-agent is the team-facing coordinator.

Responsibilities:

  • monitor inbound requests (Slack; email only when experimental mode is enabled)
  • create and manage todos
  • select target repo(s)
  • spawn and supervise dev-agent sessions
  • relay progress/results to users
  • enforce operational guardrails (cleanup, escalation)

It should remain lightweight on coding itself and focus on orchestration quality.

Dev-agent

The dev-agent is a coding worker launched in a dedicated git worktree for each task.

Responsibilities:

  • read project guidance (CODEX.md, AGENTS.md, CLAUDE.md as available)
  • implement requested changes
  • run tests/build checks
  • open PR and monitor CI
  • fix failures and address review comments
  • report completion details back to control-agent

Each dev-agent is task-scoped and should exit when work is done.

Sentry-agent

The sentry-agent handles incident-oriented analysis.

Responsibilities:

  • investigate Sentry issues by issue ID
  • summarize likely impact and root cause
  • provide actionable recommendations for fixes
  • hand off coding tasks to dev-agent through control-agent

Sentry-agent and future role agents (SRE/QA/etc.) are managed as subagent packages under pi/subagents/<id>/ and deployed to ~/.pi/agent/subagents/. Use the subagent_manage tool (or baudbot subagents ...) to install/enable/start/stop/reconcile.

Session model

  • Control and sentry sessions are long-lived.
  • Dev sessions are ephemeral and tied to todos.
  • Session-control sockets allow inter-agent messaging (send_to_session).
  • Naming conventions encode role and task context (for observability and cleanup).

Concurrency

Baudbot limits concurrent dev agents to keep resource usage predictable and avoid context thrash.

Use this model when scaling:

  • keep a fixed max active dev-agent count
  • queue additional tasks
  • prioritize by urgency or user/channel policy

Communication contract

  • Users talk to control-agent
  • Dev-agent reports to control-agent
  • Control-agent reports back to users

This keeps the external interface stable while allowing internal execution strategies to evolve.

For flow details, see team-workflow.md.