Optimization¶
Optimization is optional. Candidate HarnessSpecs and skills are staged for review. Adoption or promotion requires an explicit user action.
| Integration | Dependency | Use | Additional requirement | Detailed guide |
|---|---|---|---|---|
| GEPA | uv tool install "superqode[optimization]" | Reflective optimization for skills and harness artifacts | A reflection model and evaluation tasks | Optimization Story |
| GEPA Omni | superqode[optimization] plus a GEPA build containing the Omni API | Run GEPA, AutoResearch, and GEPA meta-harness exploration, then continue from the winner | Claude CLI for the AutoResearch and meta-harness explorers | Harness Optimization |
| AutoResearch | Installed through the GEPA Omni path | Agent-driven candidate editing and evaluation | Authenticated Claude CLI | Skill Optimization |
| GEPA meta-harness | Installed through the GEPA Omni path | Maintain and evaluate a candidate frontier | Authenticated Claude CLI | Skill Optimization |
| Superagentic MetaHarness | Install separately with uv tool install superagentic-metaharness | Export and optimize a HarnessSpec project | Selected MetaHarness backend | Harness Optimization |
| SkillOpt | Included coordinator; GEPA engines use superqode[optimization] | Optimize and stage markdown skills | Held-in and held-out task sets | Skill Optimization |
For GEPA Omni while its required API is newer than the tagged package, install GEPA from its current source branch in the same environment:
uv tool install "superqode[optimization]" \
--with "gepa @ git+https://github.com/gepa-ai/gepa.git"
Run the bounded smoke test before increasing the evaluation budget:
scripts/run_tiny_omni_experiment.sh \
--spec examples/harnesses/omni-tiny-local.yaml
Review the recorded GEPA Omni experiment before increasing the evaluation or token budget.