Skip to main content

Capability Evolution Engine

src/scheduler/capability_evolution.py. Discovers new agent capabilities from successful execution traces and registers them as named, reusable patterns.

Concept​

A "capability" is a named, parameterized procedure that the agent has learned to perform reliably. Examples:

  • "summarize a long article" → web_search → fetch_url → render_report
  • "back up a directory" → shell tar → file_ops move
  • "update a spreadsheet" → http_request → integration_skill

Capabilities are higher-level than skills (which are atomic) and persist as Capability rows in Postgres.

Detection​

The job runs hourly. It scans recent successful execution traces in execution_reflection and audit_log, looking for trace patterns where:

  • The same skill sequence repeats across goals (>= MIN_OCCURRENCES).
  • The sequence has a clear input-output signature.
  • The success rate is >= MIN_SUCCESS_RATE (default 0.8).

Storage​

Capability(
    name,                      -- generated slug
    trigger_keywords,          -- JSON array
    steps,                     -- ordered (skill, parameter mapping)
    success_count,             -- runs that completed
    failure_count,             -- runs that failed
    source_trace_ids,          -- audit log references
    created_at
)

Use during planning​

PlanGenerator includes the registered capability list in its skill catalog under "Higher-level capabilities". When the operator's request matches a capability's trigger keywords, the planner can call the capability as a single step instead of expanding the full skill sequence — saving planning tokens.

Operator review​

Generated capabilities surface at /skill-evolution (shared page with composite skills). The operator can:

  • Inspect the steps
  • Test by triggering manually
  • Approve → promotes to first-class capability used by the planner
  • Reject → marks declined; will not regenerate from same pattern within 7 days

Difference from Skill Evolution​

Both engines produce reusable abstractions, but at different levels:

Skill EvolutionCapability Evolution
OutputComposite Python skill in /data/skills/Capability row in Postgres
GranularityReplaces a multi-skill chain with one Python callHints to the planner; still expands at execution
RiskGenerated code runs in-processNo code; metadata only
Operator reviewAt /skill-evolutionAt /skill-evolution
Default stateInactive until approvedInactive until approved

Disabling​

The engine is opt-in (not auto-registered in main.py by default). To register, patch main.py:

from src.scheduler.capability_evolution import CapabilityEvolutionJob
scheduler.register("capability_evolution", 3600, CapabilityEvolutionJob())

Or feature-flag via /config.

See also​