Agentic Harness

•
10
•

Orchestrates auditable, cost-controlled LLM workflows by compiling declarative YAML playbooks into a team of AI specialists.

Agentic Harness is an open-source Python framework designed to build reliable, transparent, and cost-efficient AI agents for complex workflows. It stands out by offering unparalleled transparency through detailed markdown audit logs for every prompt, decision, and cost. It implements robust hierarchical budget enforcement with circuit breakers to prevent unexpected expenses and ensures vendor neutrality by supporting a wide array of LLMs, defaulting to local options like Ollama. This tool is ideal for backend engineers needing production-grade agents with compliance features, ML engineers requiring reproducible benchmarks, researchers conducting replayable experiments, and DevOps teams seeking seamless MCP server integration.

Características Principales

01Leverage 12 pre-built specialist agents (e.g., planner, coder, reviewer) and a curated playbook library to accelerate development.
02Define complex LLM workflows with declarative YAML playbooks, supporting conditional and parallel execution.
03Implement hierarchical budget enforcement and temporal circuit breakers for granular cost control and automatic model fallback.
04Generate comprehensive markdown audit logs and structured event logs for full transparency, debugging, and reproducibility.
057 GitHub stars
06Integrate with various LLM providers (Ollama, Anthropic, OpenAI, etc.), defaulting to local models for vendor neutrality and cost efficiency.

Casos de Uso

01Build production-ready AI agents for backend systems, ensuring compliance, budget control, and full audit trails.
02Integrate AI agent capabilities with development environments like Claude Desktop or Cursor via its native MCP server.
03Conduct reproducible benchmarks and experiments across different LLM providers for machine learning engineering and research.