> ## Documentation Index
> Fetch the complete documentation index at: https://docs.openaeon.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Aggressive Autopilot

Aggressive Autopilot is the zero-confirm execution mode for Cognitive Task OS. It
is designed for unattended task convergence: once a task is submitted, OpenAEON
can plan, execute, delegate, write back, retry, reflect, and recover without
waiting for manual approval at each phase.

## Core Capabilities

### Zero-confirm execution

When autopilot is enabled, tasks move from `PLAN` into `EXECUTE` automatically
after planning. The runtime uses legal state-machine transitions instead of
skipping directly across lifecycle phases.

### Queue autonomy

The ready queue filters blocked dependencies, releases expired claims, waits for
future `nextRetryAt` backoff windows, and can speculatively enqueue nodes whose
dependencies are still `in_progress` when speculative dispatch is enabled.

### Delegation recovery

Delegated nodes remain `in_progress` after a subagent accepts work. Runtime
summaries expose `delegations.active` and `delegations.overdue`, so the UI or
API can distinguish work that is still running from work that needs automatic
recovery.

If a delegated lease expires, OpenAEON recovers the node back to `todo` and
re-enqueues it. Ordinary dispatcher and subagent failures retry with exponential
backoff before escalating to reflection or final failure.

### Closed-loop completion

Once executable nodes are complete, the runtime can advance through `VERIFY`,
`REFLECT`, and `DONE`. Root nodes no longer block completion when all executable
children have finished; single-node tasks still fall back to the root node.

### Recursive subagent delegation

Cognitive subagents can iterate locally and spawn descendant workers when depth
policy allows. Descendants receive `parentCognitiveTask` context rather than the
parent writeback link, so only the owning subagent writes final results back to
the Cognitive Task OS runtime.

## Manual Force Start

Operators can still force a specific node to start when they intentionally want
to bypass dependency or backoff checks. The Gateway exposes this through
`cognitive.runtime.force_start`, and the runtime dispatches the requested node
rather than falling back to a generic ready-queue scan.

Use force start sparingly. It is intended for operator-directed recovery, not for
bypassing security, permission, or invariant checks.

## Runtime Signals

* **Lifecycle**: `INIT` to `PLAN` to `EXECUTE` to `VERIFY` to `REFLECT` to `DONE`
* **Queue**: pending and claimed node counts
* **Retries**: total retries, pending backoff, and exhausted retries
* **Delegations**: active and overdue delegated nodes
* **Checkpoint**: last run id and run count
* **Replay**: replay cursor for audit and inspection

## Related APIs

* `cognitive.task.submit`
* `cognitive.runtime.dispatch`
* `cognitive.runtime.force_start`
* `cognitive.runtime.status`
* `cognitive.task.replay`
* `cognitive.task.trajectory`