evalstate/fast-agent

What it solves

fast-agent is a framework for building and interacting with sophisticated multimodal AI agents and workflows. It simplifies the process of creating coding agents, development toolkits, and evaluation platforms by providing a declarative syntax for composing prompts and connecting to Model Context Protocol (MCP) servers.

How it works

The framework uses a Python-based declarative approach where agents are defined using decorators (e.g., @fast.agent). It allows developers to chain agents together into sequences, create parallel execution paths, or use advanced patterns like Evaluator-Optimizer loops and Orchestrators. It integrates deeply with MCP, supporting both STDIO and Streamable HTTP transports, including OAuth for secure server connections. It also features a CLI-first design with an interactive TUI-style terminal for live streaming responses.

Who it’s for

It is designed for developers building AI agent applications, those needing a flexible toolkit for coding agents, and teams implementing complex LLM workflows that require structured orchestration and tool use.

Highlights

• Comprehensive MCP Support: End-to-end tested support for MCP features, including Sampling, Elicitations, and Streamable HTTP transport diagnostics.
• Diverse Workflow Patterns: Built-in support for Chaining, Parallel fan-out/fan-in, Evaluator-Optimizer loops, Routers, and Orchestrators.
• MAKER (K-voting): Implements a voting mechanism to reduce errors in long chains of simple steps by sampling responses repeatedly.
• Agents as Tools: Allows complex tasks to be broken into subtasks where a main agent calls other agents as tools for routing and parallelization.
• Broad Model Compatibility: Native support for Anthropic, OpenAI, Google, Azure, Ollama, and Deepseek via TensorZero.
• Human-in-the-loop: Agents can be configured to request human input for additional context during task completion.