Agent2 v0.4.0

Agent2 v0.4.0 — Brain Clone Pipeline Upgrade

Single source of truth, operational learnings, and repo hygiene.

Added

• Operational learnings system (shared/learnings.py): auto-logs low-confidence results, clarification patterns, and rejections after each agent run. The 3 most recent learnings are injected into the next run's prompt context.
• /agent-qa skill for end-to-end testing of deployed agents with schema validation and health reporting
• Skill routing rules in CLAUDE.md for proactive skill invocation across all AI coding tools
• scripts/sync-skills.sh to keep all 5 host directories (.claude, .agents, .codex, .gemini, .github) in sync from a single canonical source
• Brain Clone SKILL.md upgrades: anti-sycophancy rules, forcing questions with pushback pattern tables, Phase 2.5 (Premise Challenge), Phase 6.5 (Scope Decision), Phase 7.5 (Architecture Review), 8-dimension agent quality rating (/80), agent slop detection (10 anti-patterns), interview session persistence, and completion status protocol

Changed

• CLI interview prompt now loads dynamically from SKILL.md instead of maintaining a hardcoded copy — SKILL.md is the single source of truth for interview phases, anti-sycophancy rules, and forcing questions
• Phase labels updated from "X/6" to "X/8" including half-phases
• _detect_phase and _detect_phase_number unified into shared keyword table
• Generator produces 5 tests (complete, clarification, rejected, before_run, resume) instead of 1
• Generator mock_result now covers all 3 outcomes including rejected
• Generator before_run always sets _instructions for learnings injection
• Generator config.yaml adds knowledge_mcp capability when collections exist
• Richer Promptfoo eval datasets generated from all example cases

Fixed

• Learnings read AND write gated behind AGENT2_DISABLE_LEARNINGS env var to prevent test pollution
• Skills synced across all host directories (previously .codex/.gemini/.github were missing skills)
• Stale "v0.3" references in README and docs
• Removed 12 tracked .playwright-mcp/ session logs, unreferenced logo (615KB), and extended .gitignore

Agent2 v0.3.0

Agent2 v0.3.0 — CLI Onboarding & Brain Clone Harness

Added

• First-class agent2 CLI with setup, onboard, doctor, list, run, serve, and publish-check commands
• Static installer scripts for macOS/Linux and Windows
• Agentic Brain Clone onboarding harness — LLM-powered adaptive interview that extracts domain expertise and generates a complete agent
• Step-by-step questionary setup wizard with model selection, profile choice, and Docker configuration
• Textual TUI alternative for fullscreen onboarding
• Central agent2.yaml framework configuration for default model, provider policy, stack profile, telemetry, and ports
• ASCII art banner in installer
• TTY detection for piped install with graceful fallback

Changed

• Normal agent configs now use model: "" and resolve through global framework config
• Docker agent images copy agent2.yaml so global config is available in containers
• Model resolution order: explicit runtime arg > agent config > agent2.yaml > env fallback

Agent2 v0.2.0

Highlights

• Adds the full Brain Clone flagship agent: procurement-compliance-officer.
• Adds public market and interview example agents.
• Updates Brain Clone skills around the Sachbearbeiter Chain-of-Thought pattern.
• Adds Agent2 onboarding docs for AI coding agents via AGENTS.md, CLAUDE.md, llms.txt, and llms-full.txt.
• Adds per-run _toolsets support from before_run() in the API runtime.
• Adds procurement knowledge books and Promptfoo evals.

Validation

• GitHub CI on main passed: Python 3.12, Python 3.13, docker-verify.
• Local validation passed: uv run pytest tests/ -v (49 passed), docker compose config --quiet.

Safety

• Release scope was scanned for secrets, private keys, tokens, and local absolute paths.
• Market agents use public read-only APIs; trading actions are paper/dry-run only.

Agent2 v0.1.0

Agent2 v0.1.0

The production runtime for AI agents. Schema in, API out.

What's included

• create_agent() + create_app() — two calls to go from schema to production API
• Typed outputs via Pydantic output_type with automatic retries
• Sync and async task execution
• Pause/resume with serialized message history
• Human-in-the-loop approval workflows
• Provider routing for prompt cache optimization
• Tool interception and collection scoping
• Knowledge search via R2R + FastMCP
• Mock mode for development without LLM keys
• RFC 7807 error responses
• Bearer token auth with rate limiting
• 39 unit tests
• Built-in skills for Claude Code, Codex, Gemini CLI, and Copilot
• 7 example agents + framework demos
• Docker Compose with default and full stack profiles

Quick start

git clone https://github.com/duozokker/agent2.git
cd agent2
cp .env.example .env
uv sync --extra dev
uv run pytest tests/ -v
docker compose up -d

Links

• Website: https://getagent2.dev
• Docs: See docs/ directory
• Comparison: When to use Agent2

MIT Licensed. Built by Artesiana.