Run autonomous workers

runtime worker is a foreground supervisor loop. It listens for eligible invocations, claims one transactionally, launches the configured provider, publishes the result, and completes or waits the invocation with evidence.

Default adapters

Use the direct adapters unless you need a specific alternative:

agent-comms --actor <agent-id> runtime worker \
  --id <runtime-id> \
  --adapter codex \
  --executable /usr/local/bin/codex \
  --codex-sandbox workspace-write \
  --execution-timeout 30m

agent-comms --actor <agent-id> runtime worker \
  --id <runtime-id> \
  --adapter claude \
  --executable /usr/local/bin/claude \
  --claude-permission-mode acceptEdits \
  --claude-max-budget-usd 1 \
  --execution-timeout 30m

OpenCode uses --adapter opencode. Its session continuity is stored in a local runtime cache because OpenCode mints non-UUID session IDs.

Adapter choices

• claude, codex, opencode: proven direct CLI execution.
• claude-live, codex-live: persistent provider processes with the broker event viewer live attach --provider claude|codex --runtime <runtime-id>.
• opencode-live: a worker-owned native server on an assigned loopback port. Copy its reported opencode attach command; the AGC broker does not have an OpenCode attach provider. The server stays alive across turns while the worker runs and stops on shutdown or one-shot exit. Existing shared native servers remain untouched.

OpenCode live workers reset only their own instance before a turn, preserve native deny rules, and verify restrictive policy before prompting. Read requests use the existing read policy, edits are allowed only in acceptEdits, and governed operations remain denied. Only an exact tool-disabled preparation message is removed; existing conversation content stays intact. Failed preparation or unknown external rule changes prevent execution. Watch through native attach without concurrent interactive prompts or policy edits during a managed turn.

A session with pending native Undo/revert state is refused before preparation: OpenCode would otherwise discard the undone history when accepting even a no-reply message. Restore or resolve that state in the native client before retrying; the worker does not discard or automatically clear it.

Turn preparation waits for the native event stream to become ready and confirm completed disposal, not just an HTTP acknowledgment. OpenCode 1.18.33’s heartbeat can add approximately ten seconds to that readiness wait. The worker reports this preparation stage; a missing completion signal stops execution safely.

Use the normal full native opencode attach TUI. OpenCode 1.18.33’s optional --mini client stops watching after per-turn instance reset; this provider limitation does not affect the supported full-TUI watcher. Copy a new reported endpoint after a worker restart.

Leave --session-id unset for OpenCode live: native IDs are non-UUID, while the public worker flag still validates UUIDs. Runtime-owned session and recovery records under the user configuration’s sessions directory provide automatic continuity. A failed resume does not silently switch to a fresh conversation.

Codex live applies --codex-sandbox and --codex-add-dir, but uses the provider’s normal user configuration. --codex-ignore-user-config is rejected before launch for codex-live and codex-acp; use --adapter codex for exec-based runs that require user MCP/tool configuration isolation.

• claude-acp, codex-acp, opencode-acp: Agent Client Protocol integrations with provider-specific permission limits.

Claude and Codex can bind a valid existing conversation with --session-id. Provider rules differ: Claude can create a caller-chosen UUID; Codex normally resumes an ID it previously minted. Never process an interactive turn in the same conversation while its worker is active.

live tail is removed. The surviving live serve / live attach path subscribes to a broker-managed runtime, not an arbitrary session log. It is distinct from runtime interactive-serve, which wraps a provider’s native terminal UI for host-local delivery.

Codex and Claude managed live workers route by stored project ID plus runtime ID. The same runtime ID can be used in independent projects sharing a broker. Run live attach in that project’s root, use --project <root>, or copy the worker’s printed --project-id <stored-id> command when viewing elsewhere. A missing scoped runtime errors; it never falls back to a host-wide runtime. For legacy/manual raw broker registrations, use live attach --unscoped. This flag is mutually exclusive with --project-id. Outside a project, with neither flag, attachment retains legacy literal-ID behavior.

During upgrade, stop old live workers and recycle your owned brokers before starting project-scoped workers. Old bare-ID processes are not silently adopted. Logical runtime IDs and existing per-project conversation caches remain unchanged. Namespace keys are not authentication or secrets; the broker’s local-host trust assumptions still apply.

Follow-up invocations

Provider shell or MCP access is not required for one agent to request another. A worker prompt accepts one bounded AGENT_COMMS_INVOKE: {json} action. The worker validates and signs the follow-up, submits it, and records the new invocation ID. Only one follow-up is accepted per completed turn to bound fan-out.

Supervise the process

Workers remain foreground processes. Use systemd, launchd, a container runtime, or your existing supervisor for restart and shutdown policy. --once processes at most one receive attempt and is intended for tests or bounded automation—not continuous autonomy.

Permission-bypassing provider modes are rejected. Output, execution time, listen intervals, and budgets remain bounded.

Exec adapters (Claude, Codex, OpenCode and custom CLI adapters) use an invocation-owned process group on Unix or job on Windows. Cancellation terminates that owned group/job, records WAITING with a bounded failure reason, and does not publish a successful result. Captured output draining is bounded to one second after direct-process exit; a child retaining output beyond that bound is a failed invocation, not a successful empty response. Unrelated provider processes and shared live brokers are not part of this cleanup.

This is process supervision, not a security sandbox. A Unix child that deliberately leaves its process group is outside group-based termination; use the provider sandbox and your host supervisor for stronger containment. ACP and live adapters have separate lifecycle management. Windows setup fails closed if the process cannot be assigned to its owned job before resume.