| name | subconscious |
| description | Self-improvement loop — agents analyze their own performance, debate improvements, and write learnings back into memory. Runs on cron schedule. |
| when_to_use | Automatically triggered by cron. Analyzes recent agent runs, generates improvement ideas, debates them via council, writes winning direction back into agent memory. |
| allowed_tools | ["memory_search","web_search","file_write","file_read"] |
| context | inline |
Subconscious Agent — Self-Improvement Loop
What This Does
The subconscious runs periodically (default: every 6 hours) and:
- Inspects — reads recent agent runs, failures, and user feedback
- Ideates — generates 3-5 improvement candidates
- Debates — challenges each idea against hard objections (uses council)
- Synthesizes — picks the winning direction
- Persists — writes the learning back into agent memory
- Reports — sends summary to supervisor (Prime)
The Loop
Recent Runs → Evidence Gathering → Ideation (3-5 ideas)
→ Debate/Critique (council) → Synthesis (1 winner)
→ Write to Memory → Next run starts smarter
Evidence Sources
- Recent session messages (last 24h)
- Tool call success/failure rates
- User feedback (thumbs up/down)
- Escalation history
- Memory staleness
What Gets Written Back
- Winning direction → saved as agent memory (explicit certainty)
- Rejected paths → saved as agent memory (abductive certainty, low weight)
- Improvement backlog → saved as prime memory
- Run summary → saved to audit log
Configuration
schedule: cron expression (default: 0 */6 * * * — every 6 hours)
max_ideas: number of candidates to generate (default: 5)
debate_rounds: number of challenge/defense rounds (default: 2)
auto_apply: whether to apply improvements without approval (default: false)
model_ideation: model for generating ideas (default: cheap/fast)
model_debate: model for critique (default: premium)