LOGICAL LENSES / AGENT ENGINEERING 01 THE AGENT HARNESS — FIELD NOTES THE CENTRAL IDEA The harness is the control software that runs the model-driven interaction. It prepares input, calls the model, interprets output, routes proposed tool calls, returns observations, and manages whether the run continues, pauses, or ends. Implementations arrange these responsibilities differently. KEEP THE ROLES DISTINCT Model: generates responses and proposed actions. Harness: controls the interaction and routes proposals. Tool: performs an operation and returns a result. Execution environment: where the operation runs and what it can access. Application: connects the system to the user and the wider product. CONTEXT VS TASK STATE Context is the information available to the model for a particular call. Task state is the system's record of progress and results. Some task state can be selected into later context. The terms are not synonyms. FOLLOW THE EVIDENCE 1. The user asks the agent to investigate a failed deployment. 2. The status tool confirms the deployment failed. 3. Reading logs times out. No log evidence has been returned. 4. A permitted second read returns a missing-configuration startup error. 5. The model can now draft a diagnosis grounded in the returned evidence. 6. A proposed ticket-creation operation is held for user approval. 7. After approval, the tool creates the ticket and returns an identifier. THREE DIFFERENT CLAIMS Proposed: the model requested an operation. Attempted: the system dispatched the operation. Confirmed: a returned result provides evidence of the outcome. A lost response does not prove a remote operation failed. It may have completed even when the current run cannot confirm the result. THE LIMIT A harness does not guarantee correct model responses. The execution trace helps check the model's claim against the actions and observations. NEXT: AGENT CONTROL LOOPS After an observation, what determines whether the system continues, tries another step, pauses, or stops? Channel: https://www.youtube.com/@LogicalLenses