Skip to content

README.md

Latest commit

 

History

History
92 lines (75 loc) · 4.4 KB

README.md

File metadata and controls

92 lines (75 loc) · 4.4 KB

AIDE Evolution Engine (v2.1 — Distributed Agent-Native Scheduling)

The Evolution Engine is AIDE v2.0's autonomous self-evolution system. As of 2026-05-04 (RFC-0003) the heavy CI-cron pipeline has been retired. Each participating agent now runs on its own platform's scheduler and opens PRs into this repository directly.

Architecture

                      Weekly Intel Loop (sensor, keyless)
                      Mon 00:00 UTC — vendors, HN, blogs, benchmarks
                                     │
                                     ▼ commits weekly digest, surfaces signals
   Distributed scheduled agents (per-platform schedulers, no shared CI):

   Mon 00:00 UTC  Claude Code on web   ──>  curates digest, opens PR with
                  (Anthropic side)          AIDE-relevance commentary +
                                            principle re-tuning candidates
   Mon 00:00 UTC  Codex app schedule   ──>  adversarial review of the
                  (OpenAI side)             Claude PR; blocks or approves;
                                            posts critique
                                     │
                                     ▼  consensus (≥2 different-vendor agents)
                  Axiom Gate CI       ──>  enforces A1-A5; merges on pass

The Weekly Intel Loop (evolution/intel/, workflow aide-weekly-intel.yml) remains the cheap, keyless mechanical sense pass. See evolution/intel/README.md.

The per-platform scheduled agents replace the previous CI cron because:

  • Platform schedulers (Claude Code on web, Codex app) handle auth, retries, tool access, and budgets natively — no GitHub Actions secrets to manage.
  • Each agent stays in its native environment (no CLI bridge layer).
  • A schedule outage on one platform doesn't block the others.
  • Gemini is currently out of the loop (no scheduler support yet); axioms A2 / A4 were updated to require ≥2 different-vendor agents instead of 3, with an explicit tie-breaker rule (1:1 split blocks).

Directory Structure

evolution/
  scripts/intel/        # Weekly Intel Loop — external signal scanners
    fetch_vendor_releases.py  # Anthropic / OpenAI / Google releases
    fetch_social_signals.py   # HN, tech blogs, X, Threads
    fetch_benchmarks.py       # SWE-bench, Terminal-bench, WebArena, SWE-rebench
    compile_weekly_digest.py  # Merge + dispatch decision + source health
  benchmarks/           # Reserved for benchmark snapshots fetched by agents
  deliberation/         # Reserved for agent deliberation artifacts (PR-attached)
  sandbox/              # Reserved for empirical validation results
  history/              # Evolution audit trail (permanent record)
  intel/                # Weekly intel digests (permanent MDs + transient YAMLs)

The benchmarks/, deliberation/, sandbox/ directories are now populated by the scheduled agents themselves on a per-PR basis rather than by a CI pipeline.

Triggering

  • Weekly machine sense: aide-weekly-intel.yml runs every Monday 00:00 UTC.
  • Curated proposal: Claude Code on web is scheduled to run every Monday and opens a PR with the curated digest and any principle re-tuning candidates.
  • Adversarial review: Codex app (OpenAI) is scheduled to review the Claude PR within 24 hours. Approval from the Codex agent satisfies A4 (≥2 different- vendor consensus).
  • Manual trigger: gh workflow run aide-weekly-intel.yml.

Required Secrets

Repository-side: none required for the new pipeline. Each scheduled agent runs in its own platform with its own auth.

CI-side (the workflows that remain in this repo):

Workflow Secret Purpose
aide-weekly-intel.yml none Reads only public RSS / APIs
axiom-gate.yml none Static checks on PR metadata

Axiom Compliance

Every evolution PR verifies compliance with the 5 Immutable Axioms:

  • A1 Reversibility: All changes are git-backed and revertable.
  • A2 Adversarial Separation: ≥2 different-vendor agents participate (e.g., Claude proposes, Codex reviews).
  • A3 Empiricism: Changes require quantitative evidence (benchmark deltas, repo metrics) — agent reasoning alone is insufficient.
  • A4 No Single Authority: Consensus from ≥2 independent agents required; a 1:1 split blocks the change (no auto-resolution).
  • A5 Self-Observability: Full audit trail in evolution/history/ and the weekly digest under evolution/intel/.