Summary

AgentOps provides a DevOps layer for coding agents, addressing their forgetfulness and inconsistency by enabling compounded flow, feedback, and memory across sessions. It transforms agents into reliable contributors by building durable, repo-native institutional knowledge, benefiting users seeking to maximize agent productivity and code quality through persistent learning and validation.

How It Works

AgentOps centers on three capabilities: judgment validation (e.g., /pre-mortem challenges plans, /council validates code), durable learning (via repo-native memory in .agents/), and loop closure (completed work enhances future tasks). Its "Knowledge Flywheel" captures lessons, promotes them (lesson -> pattern -> rule), and injects them into future sessions. Memory is stored in plain-text .agents/ files, allowing diffing, PR review, and grep-based search, offering an auditable, persistent alternative to traditional RAG.

Quick Start & Requirements

Installation varies: Claude Code (marketplace add boshu2/agentops), Codex CLI (install-codex.sh), OpenCode (install-opencode.sh), or Skills-compatible agents (npx skills@latest add ...). The ao CLI is available via Homebrew (brew install agentops). Linux users may need bubblewrap for Codex. Key docs: docs/newcomer-guide.md, docs/SKILLS.md.

Highlighted Details

• Repo-Native Memory: .agents/ stores knowledge as plain, diffable Markdown files, integrating with version control.
• Knowledge Flywheel: Lessons compound across sessions, agents, and runtimes, turning debugging into reusable institutional knowledge.
• Multi-Model Validation: Skills like /council and /vibe use independent Claude + Codex judges for plan and code validation.
• evolve Skill: Enables unattended, fitness-scored codebase improvement loops that automatically address identified gaps.
• ao CLI: Offers terminal-based knowledge management, autonomous cycles, and control-plane workflows.

Maintenance & Community

GitHub Actions indicate active automated testing. No specific community channels or notable contributors/sponsorships are detailed in the README.

Licensing & Compatibility

Licensed under Apache-2.0, AgentOps operates locally with state on disk, not modifying source code. This ensures compatibility with closed-source projects and various environments (Claude Code, Codex CLI, Cursor, OpenCode).

Limitations & Caveats

Support for integrations like Cursor and OpenCode is experimental. Qualitative observations suggest agent effectiveness hinges on workflow, context injection, and validation, not just prompts. Raw chat history requires distillation into structured knowledge for utility.