docs: add adaptive agent selection problem document#389
docs: add adaptive agent selection problem document#389fullsend-ai-coder[bot] wants to merge 1 commit into
Conversation
Add a new problem document exploring how fullsend can learn which agent/team/workflow configurations are effective for which classes of problems using evolutionary algorithms (GA) and Thompson Sampling. The document covers: - Three optimization surfaces: team selection, agent configuration, and workflow shape - Candidate approaches: Thompson Sampling for discrete selection, GA for structured workflow evolution, and a hybrid - Fitness function design with gameability analysis and difficulty normalization - Context granularity trade-offs for bucketing fitness data - Cold-start strategies and similarity-based fallback - Shadow mode as a prerequisite for live deployment - Safety constraints preserving deterministic safety model - Storage and observability requirements - Minimum viable experiment design Also links the new document from README.md. Note: make lint could not run due to sandbox network restrictions (gitleaks hook download blocked). Pre-commit infrastructure error (exit 3), not a code issue. Closes #350
Site previewPreview: https://208bfbd2-site.fullsend-ai.workers.dev Commit: |
![fullsend-ai-review[bot]](https://avatars.githubusercontent.com/u/271249281?s=60&v=4){kind=small}
There was a problem hiding this comment.
Review: #389
Head SHA: 7bd494f
Timestamp: 2026-04-24T00:00:00Z
Outcome: approve
Summary
This PR adds a well-structured problem document exploring adaptive agent selection using evolutionary algorithms and Thompson Sampling, along with a README link. The document follows repo conventions — it presents multiple approaches with trade-offs, includes an extensive Open Questions section, and correctly cross-references five existing problem documents (all verified to exist). The change is tightly scoped to what issue #350 authorized. No code changes, no security implications. All internal links are valid. The document's treatment of safety constraints (deterministic safety preservation, self-referential fitness avoidance, optimization pressure on safety paths) demonstrates appropriate awareness of the repo's threat model.
Findings
Info
- [style] PR body — The PR description notes "make lint could not run due to sandbox network restrictions." CLAUDE.md requires running
make lintbefore submission. Since this is a documentation-only change (Markdown files), lint risk is negligible, and the post-script verification section confirms pre-commit hooks passed on the runner. Non-blocking.
Footer
Outcome: approve
This review applies to SHA 7bd494f8604fc065bab1cec8fce80988a25ce065. Any push to the PR head clears this review and requires a new evaluation.
Add a new problem document exploring how fullsend can learn which agent/team/workflow configurations are effective for which classes of problems using evolutionary algorithms (GA) and Thompson Sampling.
The document covers:
and workflow shape
GA for structured workflow evolution, and a hybrid
normalization
Also links the new document from README.md.
Note: make lint could not run due to sandbox network restrictions (gitleaks hook download blocked). Pre-commit infrastructure error (exit 3), not a code issue.
Closes #350
Changed files
README.mddocs/problems/adaptive-agent-selection.mdCloses #350
Post-script verification
agent/350-adaptive-agent-selection)c1e8844bcd6b9900c031eb7b11e28ec4d576717d..HEAD)Created by fullsend code agent