harness
The 'cockpit' bolted around an AI model so it can actually do things — click tools, remember context, take steps — instead of just chatting.
The runtime/scaffolding around an LLM: tool dispatch, context management, and the agent loop (plan → act → observe). Colloquially 'MCP for agents'.
The agent execution environment hosting an LLM: tool/permission brokering, context-window and memory management, the perceive-decide-act control loop, and integration wiring (e.g. MCP servers). Distinct from the model weights.
evolution
- 2022-11 · historyLaunch of ChatGPT
The release of ChatGPT demonstrated the potential for LLMs to follow instructions, sparking immediate interest in wrapping them with external logic.
- 2023-01 · historyLangChain Release
Harrison Chase released LangChain, the first major framework specifically designed to chain LLM calls with external data sources and tools.
- 2023-03 · historyAutoGPT Emergence
AutoGPT popularized the concept of autonomous agent loops, where an LLM continuously prompts itself to complete multi-step tasks.
- 2023-11 · historyOpenAI Assistants API
OpenAI integrated 'harness' capabilities directly into their platform, providing native support for code execution, file retrieval, and persistent threads.