Alex Xu
How does an agent harness work? An LLM can reason and generate text, but that alone can be unreliable. The reliability usually comes from the harness, the layer around the LLM shown in the diagram below. 1. The user sends a query into the harness and the context builder converts the query into a prompt using its knowledge base. 2. The LLM reasons over the provided prompt and proposes some actions (functions to call) 3. The policy gate checks actions and determines whether they should be allowed or blocked. 4. The runtime first executes the tools and sends the results back to the LLM. 5. Finally, the verify step runs tests on the result and includes the feedback in the context for future iterations. 6. Once the verify step passes, the harness shows the result to the user. Aside from these, two other layers exist and are shared between all stages. The observability layer tracks things like metrics, and errors. The constraints layer sets the rules and permissions so the system can enforce them. Where do you think most agent failures happen, in context, tools, constraints, or verification?