Open specifications and SDKs for AI agent interoperability.
15 protocol specifications — communication, identity, skills, trust, economics, compliance. CC BY 4.0. Single portal at workswithagents.dev.
See workswithagents.dev/specs/ for the full index.
- Python:
pip install workswithagents - TypeScript:
npm install @workswithagents/agent-foundry - Source: github.com/workswithagents/works-with-agents
The 10-pattern agent methodology is the core of Works With Agents. These aren't theoretical — they emerged from a 3-day experiment where AI agents scaffolded 111 SPFx web parts and 5 backend services autonomously.
Key patterns: bootstrapping, skill composition, persistent memory, decision protocols, tool composition, orchestration, pipelines, resilience, verification, and compounding.
We test LLMs nightly for both code quality and agent readiness (tool-calling reliability). The gap is massive — models that score 90%+ on code quality often fail at agent tasks.
What we test:
- Code quality: 10 real coding tasks — build, deploy, fix
- Agent readiness: 6 tool-calling tests — single-tool, multi-tool, required mode, false positives, multi-turn, argument correctness
- Local models: Gemma 4 E4B, Qwen 3, SmolLM3, Phi-4-mini, Bonsai 1-bit
- Cloud models: Claude Opus, GPT-5, Gemini, DeepSeek, Grok, and 90+ more
Vilius Vystartas — Technical Pelin Kayhan — Business & Compliance