/darwin

Subsystems

Darwin's 7 internal subsystems that compose the SENSE→ASSESS→EVOLVE→VERIFY→PERSIST framework.

Prior art Darwin's mutation operators borrow from (2026-05 refresh)

• AlphaEvolve (Google DeepMind, Gemini-powered evolutionary coding agent, arXiv 2506.13131): combines a MAP-Elites-inspired strategy with an island-based population model to maintain both performance and diversity. Notably discovered a 48-scalar-multiplication procedure for 4×4 complex matrix multiplication — the first improvement over Strassen's algorithm in 56 years — and was applied internally at Google to data-center scheduling and LLM training acceleration. Darwin's Affinity Evolver and Fitness Scorer treat MAP-Elites-style diversity preservation as the reference pattern: never collapse Dynamic AFFINITY into a single dominant agent stack when phase signals are mixed; preserve niche agents the way MAP-Elites preserves behavioral cells. Source: https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/.
• Anthropic Dreaming (research preview, 2026-05-06): between-session memory consolidation that extracts patterns from past sessions, surfaces recurring mistakes, and keeps shared team preferences fresh. Darwin's Journal Synthesizer is the in-ecosystem analogue — both produce reusable patterns; both must be governed by a review-before-merge default to prevent silent memory drift. Source: https://claude.com/blog/new-in-claude-managed-agents.
• Anthropic Outcomes (public beta, 2026-05-06): rubric-driven success criteria with measured +up-to-10 pp task-success lift. Darwin's Trigger Engine ET-02 (Quality Plateau) should now emit Outcomes-style rubrics as the proposed remediation, not just generic "improvement chain" text.

1. Lifecycle Detector

Determines the current project phase from environmental signals. Runs automatically at the start of every Darwin invocation and on ET-07 trigger.

Process:

1. Collect git metrics (commit frequency, file types, branch patterns)
2. Analyze file structure (test presence, CI configs, deploy configs)
3. Score each phase (0.0-1.0) based on signal matches
4. Select highest-scoring phase (or mixed if <0.60)
5. Compare with previous detection for transition events

2. Trigger Engine

Evaluates trigger conditions (ET-01 through ET-08 as the base set; ET-09 Official Spec Conformance drift in official-fitness-criteria.md; ET-10 Framework End-of-Life drift added 2026-05 — see evolution-actions.md) and fires appropriate evolution actions.

Process:

1. Check each trigger condition against current state
2. Prioritize triggered actions (ET-05 emergency > others)
3. Execute actions in priority order
4. Log trigger events to ECOSYSTEM.md

3. Journal Synthesizer

Analyzes .agents/*.md journals to extract cross-cutting patterns.

Process:

1. Scan all journal files for entries with reusable: true tag or high-value patterns
2. Cluster related entries across agents
3. Generate Pattern Cards: {pattern_id, source_agents, insight, apply_when, confidence}
4. Store in ECOSYSTEM.md Cross-Agent Discoveries section

4. Affinity Evolver

Computes Dynamic AFFINITY overrides based on lifecycle, usage, and feedback.

Process:

1. Get current lifecycle phase
2. Map phase to dominant agent profiles (see Lifecycle table)
3. Cross-reference with actual usage from PROJECT.md
4. Apply feedback modifiers from Reverse Feedback
5. Output override entries for ECOSYSTEM.md

5. Discovery Propagator

Creates knowledge transfer briefs when one agent's finding benefits others.

Brief format: see evolution-actions.md § Discovery Propagation (canonical DISCOVERY_BRIEF schema).

6. Staleness Detector

Evaluates each agent's ongoing relevance to the current project.

Process:

1. Calculate RS for every agent in the ecosystem
2. Flag agents with RS <40 as Dormant
3. Flag agents with RS <20 as Sunset candidates
4. Cross-reference with lifecycle phase (some dormancy is expected)
5. Generate Staleness Report with recommended actions

7. Fitness Scorer

Calculates the Ecosystem Fitness Score (EFS) across 5 dimensions.

Process:

1. Calculate each dimension score (0-100)
2. Apply weights: Coverage(25%) + Coherence(20%) + Activity(20%) + Quality(20%) + Adaptability(15%)
3. Compute trend by comparing with previous EFS (up/down/stable)
4. Assign grade (S/A/B/C/D/F)
5. Generate dashboard summary