Skip to content
github-actions[bot] edited this page Sep 17, 2026 · 8 revisions

conductor

Event-driven agent orchestration for your Paseo daemon: connect to services once, declare where agents run, and wire events to work — agents, commands, inline code, and any connector's verbs, crossing service boundaries freely.

connectors:
  gh:        { use: github, app: {…}, me: { logins: [your-login] }, repos: ["your-org/*"] }
  slack-ops: { use: slack, app_token: ${SLACK_APP_TOKEN}, bot_token: ${SLACK_BOT_TOKEN} }

runtimes:
  paseo: { use: paseo, default: true }

x-templates:
  fixer: &fixer { type: agent, workspace: worktree }

triggers:
  - on: gh.merge_conflict
    steps:
      - { <<: *fixer, id: fix, prompt: "Resolve the conflict on {{.repo}}#{{.pr}}." }
    hooks:
      - { at: done, uses: slack-ops.post, options: { text: "resolved {{.repo}}#{{.pr}}" } }

What sets it apart

Most workflow tools wire events to API calls. conductor's steps run autonomous coding agents — they check out a repo, edit in a worktree, run commands, and open a PR — and the platform is built to direct and contain that:

  • Agent-authored workflows — an agent can plan a multi-step workflow at runtime and run it under guardrails: an allow-list of the steps, connectors, and secrets it may touch, so the plan can't reach past what you granted (Workflows).
  • Quality gates on agent output — tests, lint, or a critic agent must pass before a proposed change lands; a failure loops back a revise (Gates).
  • Governance for autonomous work — cost/token budgets (Cost-Accounting), OS-enforced isolation with an egress allow-list (Isolation), and a secret broker so agents act through conductor without seeing raw secret values (Agent-Skill).
  • Multi-agent teams (Teams), scoped memory across runs (Memory), and an outcome-learning loop (Outcomes).

Another orchestrator (n8n and the like) can also call conductor's authenticated /invoke API for the agent work it does (Callable-Service).

The model

  • Connectors — external services, each with events (on:) and verbs (uses:), self-describing schemas, and per-connector policy. Anything without a built-in type is a rest / graphql connector declared in config.
  • Runtimes + Steps — where agents run and who they are; an agent profile's session: binds one live agent per key (Steps).
  • Workflows — the trigger grammar: on / filter / steps / hooks, position-scoped context, control flow, reusable workflows, and agent-authored plans.
  • Verbs — the shared action unit, option merging, as: identity, and request-response ask (Hand-offs).
  • Code-Steps — the builtin cli engine, sandboxed engine plugins (js, go-embed, risor, lua) and host interpreters; Hosts for SSH remote execution.
  • Stores & Memory — named stores: (KV + SQL) behind the always-on kv.* / sql.* verbs, and shared agent memory with provenance and scope.
  • Grouping — debounce batching, one run per key.
  • Policy — quiet hours, concurrency, ignores, rate limits, pause labels, reply_to_bots, and the agent_authored guardrails; global / connector / trigger, most specific wins.
  • Secrets — ${ENV} as the baseline plus named vaults: (conductor / onepassword / pass / file / hashicorp), one {{ vault … }} reference syntax, tainting/redaction, and the non-interactive unlock model. Agent-Skill — how a dispatched agent reaches back into conductor (verb tools + the secret broker), gated per profile.
  • Gates — quality gates on agent output: checks against the proposed change, a bounded revise loop, escalation. Teams — one task split across planner / parallel workers / critic / reconciler.
  • Outcomes — merged / reverted / approved / rejected captured per agent action and fed back into memory, workflow health, and guidance; Cost-Accounting — per-run token/$ usage and hard budget caps.
  • Runs — full execution history, conductor watch live streaming, proposed-diff preview, and retry-from-step. Isolation — per-dispatch sandboxing and the egress allowlist. Binary-Data — files as content-addressed blob handles between steps and agents.
  • Authoring-Connectors — adding a connector type in-tree (the contract, the executable template, build-tag pattern).
  • Plugins — external connectors & runtimes: out-of-process plugins the daemon runs to add a type: or runtime: without recompiling, and the security model that gates them (verify, sandbox, least-privilege creds).

Learning path

Work down this list and you go from zero to the most advanced setup:

  1. Install and first trigger — Installation, then Quickstart: a credential-free config, conductor validate, and a manual run.
  2. Core concepts — Connectors (events vs verbs), Workflows (triggers, steps, hooks, context scope), Verbs (options + identity).
  3. Connect real services — GitHub App setup for GitHub, then browse the connector catalog (each with its own setup) and the built-in types in Connectors; anything not yet typed via rest/graphql in Configuration.
  4. Agents — Runtimes, Steps (profiles, checkout, guidance), Hand-offs (ask verbs), and Notifications (the conductor.* lifecycle source).
  5. Composition — reusable workflows with inputs/outputs (Workflows), YAML anchors, extends: inheritance + layered guidance (Reuse), Grouping, Code-Steps, Hosts, and file splitting (imports:, Configuration).
  6. State — stores: + the kv.*/sql.* verbs (Stores), Memory, and session affinity (Steps).
  7. Hardening — Secrets (vaults, OAuth2 logins, unlock), Policy (quiet hours, rate limits, bots), and dry-run/replay (Commands).
  8. Advanced — agent-driven workflows and policy.agent_authored (Workflows + Policy), the agent skill (Agent-Skill), self-update as a workflow (Configuration), remote fleets over SSH (Hosts).
  9. Agent quality — Gates on agent output, Teams for one big task, Outcomes closing the loop, Cost-Accounting budgets, Isolation sandboxing, Runs (history / watch / retry), Binary-Data artifacts.
  10. Growing the library — Authoring-Connectors.
  11. Coming from the legacy schema — Migration (automatic, total, fail-safe).

Pages

Setup: Installation · Quickstart · GitHub App setup ↗ · Configuration · Commands · Examples

The model: Connectors · Workflows · Reuse · Settings-and-Templating · Verbs · Code-Steps · Stores · Runtimes · Steps · Grouping · Memory · Binary-Data · Agent-Skill · Policy · Gates · Teams · Outcomes · Cost-Accounting · Secrets · Trust-and-Isolation · Hosts · Isolation

Connector references: Connectors · Authoring-Connectors · connector catalog ↗

Operations: Runs · Notifications · Hand-offs · Migration (the legacy schema still loads and auto-migrates)

Clone this wiki locally