fast-agent

fast-agent is an open-source Python harness for building AI agents that takes the unglamorous parts seriously: provider abstraction across OpenAI, Anthropic, xAI, DeepSeek, Moonshot, Google, llama.cpp, Azure, and Bedrock; native MCP client and server support (plus ACP and A2A protocols); a terminal TUI; structured outputs; prompt caching; compaction; and a privacy filter. You define agents in config or code, run workflows, and deploy them.

What sets it apart is the evaluation story: built-in benchmark harnesses with published results and methodology, batch processing, and GEPA optimization for improving agent behavior — it treats agents as something you measure, not just something you demo. It also plays well with the ecosystem: PulseMCP's own servers list fast-agent as a target framework, and it supports agent skills over MCP and the OpenAI Apps SDK.

The honest caveat: it is a framework, so you still write the agent logic — this is not a no-code builder. The docs are extensive but the surface area is large, and provider OAuth/subscription logins (Codex, Copilot) add setup wrinkles. For developers who want one harness across many models with evaluation built in, it is one of the strongest options in 2026.

1. Install astral uv, then run uvx fast-agent-mcp@latest (or pip install fast-agent-mcp).
2. Run the setup guide to configure providers (API keys, OAuth, or local models via llama.cpp).
3. Define agents in fastagent.config.yaml (or the Python API) and connect MCP servers.
4. Build and iterate in the TUI; run benchmarks to evaluate before deploying.