from ASI-Evolve669
Run an ASI-Evolve style loop: preflight, learn, design, experiment, analyze. Use when you need structured evaluator-driven iterative improvement for code, promp
Evolve provides a disciplined single-agent abstraction of an evaluator-driven search workflow (learn -> design -> experiment -> analyze). It wraps repository scripts and enforces a gated preflight, structured rounds, a cognition store for external insights, and an experiment database for traces and results. The skill is designed for safe, auditable iterative improvement on artifacts such as code, models, or prompts where each round's proposal is evaluated and recorded.
Use Evolve when you want to run controlled optimization experiments: tune a prompt, search for better code variants, or iterate on evaluation pipelines. It's appropriate when you need explicit preflight confirmation (objective, scoring, evaluator command, timeout, writable scope), reproducible database-backed runs, and careful serialization of DB operations. Do not use it to bypass preflight or to run unapproved mutating commands.
scripts/ wrappers for DB and cognition operations, run serialization, and evaluation tooling (has_scripts=true)references/ contains operating model docs, preflight guidelines, run spec schema, and architecture notes (has_references=true)Best suited to agent runtimes that can run repo scripts and manage file-scoped mutations (Codex/Copilot-style or self-hosted agent runners). It expects the agent to run CLI wrappers and maintain local run directories as the source of truth.
Evolve SDK skill provides comprehensive documentation for building sandboxed CLI agent applications in TypeScript and Python. No bundled scripts to execute — purely a reference/documentation skill. SKILL.md is well-organized with clear language detection logic, required reading sections, and an extensive topic index covering sandbox config, streaming, swarm pipelines, and observability. The self-update mechanism via git clone without confirmation is a minor security concern.
Clean documentation-focused skill. Self-update command is the only notable security finding — it clones from the official repo and copies files without explicit user confirmation, but it's a standard pattern for skill updates. No scripts to audit. Well-structured with good separation of concerns between the overview SKILL.md and detailed reference docs.
