Fellow is the system's empirical optimization engine, invoked exclusively by Mentor to determine which implementation of a skill, prompt, heuristic, or workflow actually performs best -- not which one looks best on paper. It runs controlled experiments with a fixed benchmark and compute budget, establishes a fresh baseline before testing any variant, and returns the winning result with full mutation lineage so every promotion is traceable and reversible.
Skill packages follow the agentskills.io open standard and are compatible with OpenClaw, Hermes Agent, and any agentskills.io-compliant client.
Fellow exists because "looks better on paper" is not the same as "works better in practice." When Mentor identifies a skill or heuristic that needs improvement, Fellow takes over the empirical work -- establishing a fresh baseline, generating and testing controlled variants within a constrained mutation surface, and returning the winning implementation with full lineage. Every experiment runs against a fixed benchmark and compute budget so results across different targets remain comparable and every promotion is reversible. Fellow is not user-invocable; it is called only by Mentor.
| Command | Description |
|---|---|
fellow.experiment.run |
Execute an experiment cycle from a Mentor invocation payload |
fellow.experiment.status |
Current experiment state if a cycle is in progress |
fellow.journal |
Write journal for the current run |
fellow.update |
Pull latest from GitHub source (preserves journals and data) |
fellow.init runs automatically on first invocation by Mentor and creates all required directories, config.json, and JSONL files. No manual setup is required. Fellow is purely reactive -- invoked only by Mentor, with no cron jobs or heartbeat entries.
OCAS Skills
- Mentor -- sole invoker; provides experiment programs and approves promotions
- Elephas -- stores experiment lineage and artifacts via journal signal payloads
External
- None
| Job | Mechanism | Schedule | Command |
|---|---|---|---|
fellow:update |
cron | 0 0 * * * (midnight daily) |
Self-update from GitHub source |
- Version alignment per spec-ocas-skill-publishing.md (no functional change)
- Structured entity observations in journal payloads (
entities_observed,relationships_observed,preferences_observed) user_relevancetagging on journal observations (defaultagent_onlyfor experiment entities,userfor user-preference experiments)- Elephas journal cooperation in skill cooperation section
- Added
fellow.updatecommand and midnight cron for automatic version-checked self-updates
- Routing aliases and trigger phrases
- Full experiment lifecycle implementation: baseline, variant generation, benchmark execution, metric extraction, promotion
- Invocation contract schema
- Initial release as part of the unified OCAS skill suite
Fellow is part of the OCAS Agent Suite -- a collection of interconnected skills for personal intelligence, autonomous research, and continuous self-improvement. Each skill owns a narrow responsibility and communicates with others through structured signal files, shared journals, and Chronicle, a long-term knowledge graph that accumulates verified facts over time.