Autopilot learning

agentic-flow's autopilot learning bridge (ADR-058) records each task episode and, when a task comes back, hands the agent what earlier attempts learned. In Claude Code you get the same loop with the ReasoningBank tools on the agentic-flow MCP server. Recording needs the learning or admin MCP profile (plugin setting mcp_profile); recall works under every profile.

Each iteration

1. Pick the next task (from the task list, a TODO file, or the user).
2. Recall: agentdb_pattern_search with the task text, k: 5. Read failures (onlyFailures: true) before successes when the task was attempted before.
3. Plan in one paragraph that cites what the recall showed, or says nothing matched.
4. Do the task and verify it with something checkable (tests, a build, a command's output).
5. Record: agentdb_pattern_store with success, an honest reward, the number of attempts in critique, and what to do differently.
6. Re-engage a failed task only with a changed approach; the same plan twice is a loop, not learning. After three failed attempts, stop and report.

Guard rails

• Each agent spawned through agentic_flow_agent gets read-only tools unless the plugin's allow_bypass_permissions setting is on. Leave it off outside a disposable sandbox.
• The plugin's guard refuses agentic_flow_agent tasks that match destructive or exfiltration patterns. It is a pattern list, not a sandbox: do not rephrase a refused task to get past it; ask the user.
• Do not store secrets or personal data in episodes.

Report

At the end: tasks done, tasks failed, attempts per task, and the critiques worth keeping.