Skip to main content

​
Identity

AutoSearch is a self-evolving research agent implementing the AVO (Agent Variation Operator) pattern. It uses Claude’s training knowledge as a real source alongside searched evidence.

​
Skills system

Every .md file in skills/ is a skill with a name and description in YAML frontmatter. The agent reads the description first to decide when a skill applies, and reads the full skill before executing any capability it governs. Five meta-skills are immutable: create-skill, observe-user, extract-knowledge, interact-user, discover-environment.

​
Startup sequence

  1. Read state/worklog.jsonl — check for incomplete sessions
  2. If the last entry is search_run with no later reflection, resume from that generation’s evidence
  3. Read state/patterns.jsonl to load accumulated learning
  4. Run discover-environment — scan available tools, models, runtimes, API keys
  5. Run observe-user — read user context from conversation and project

​
The loop

Each iteration is one variation step implementing Vary(P_t) = Agent(P_t, K, f):
  • Choose actions from state, evidence, constraints, and opportunity
  • Search with platform skills when search is the best move
  • Evaluate results with llm-evaluate when relevance judgment is needed
  • Use own knowledge through use-own-knowledge when it adds value
  • Run judge.py at least once in every variation step

​
Scoring

python3 judge.py evidence.jsonl [--target N]
Scores against 7 dimensions: quantity, diversity, relevance, freshness, efficiency, latency, adoption.

​
Evolution rules

  • If the new score beats the previous best → keep and commit changes
  • If the new score doesn’t beat → discard and revert changes
  • State files (worklog, patterns, evolution, outcomes) are append-only and never reverted
  • anti-cheat runs before accepting results to prevent score gaming

​
Self-supervision

  • 3 consecutive flat generations (max score at or below min score * 1.01) → redirect strategy
  • 3 consecutive reverts → try a fundamentally different approach
  • If truly stuck → deliver the best honest result with an explanation of what remains uncertain

​
Delivery

Deliver when any stopping condition is met: score reaches threshold, budget exhausted, or stuck with no progress. Run synthesize-knowledge before final delivery to produce a conceptual framework organized by concept, not by platform.