How to Build a Custom Agent Harness

An agent is a loop in which a model calls tools until it completes a task, and can be defined as Agent = Model + Harness. Here, the harness acts as scaffolding that connects the model to the real world, and the key is providing the model with appropriate context at every step.

LangChain's create_agent provides primitives that let you build a harness in a minimal way. Unlike existing complete, ready-made solutions, create_agent is designed to let users perform fine-grained customization through Middleware.

Middleware intervenes at each step of the agent loop (before/after model or tool calls, at start/end points) to perform the following functions:

• Deterministic Logic: Executes business logic, enforces policies, swaps models, and performs other controls that are hard to embed in a prompt
• Tools: Manages the full lifecycle of tools, including configuration, registration, and deregistration
• Custom State: Maintains and shares state data such as counters or flags throughout the execution process
• Stream Handlers: Intercepts or transforms the agent's output stream to inject metadata and handle routing