🔧 Harness & Loop Engineering¶
Building the production scaffolding around a model (the harness) and the iterative control flow that drives it (the loop).
Overview¶
A model on its own only turns tokens into tokens. Everything that lets it act safely in production (tools, sandboxes, permissions, verification, memory, tracing) is the harness, and the control flow that repeatedly calls the model, runs tools and decides when to stop is the loop. When teams use comparable frontier models, these two layers decide most of the reliability, cost and safety differences you see in production.
flowchart TB
subgraph H["Harness (scaffolding)"]
G["Guardrails and permissions"]
S["Execution sandbox"]
V["Verification (tests, validators, judges)"]
O["Observability (traces, cost, audit)"]
end
subgraph L["Loop (orchestration)"]
C["Assemble context"] --> M["Call model"]
M --> A["Run requested tools"]
A --> E["Observe results"]
E -->|"not done, within budget"| C
E -->|"done or budget hit"| X["Exit"]
end
H -. "wraps every step of" .-> LThe common shorthand is Agent = Model + Harness: the harness is everything except the model weights. The term was popularised in early 2026 by OpenAI's write-up on building with Codex and by Birgitta Böckeler's article on martinfowler.com, which frames the harness as guides (feedforward controls) and sensors (feedback controls). Anthropic's engineering posts ("Building effective agents", "Effective harnesses for long-running agents", "Effective context engineering for AI agents") cover the same ground with concrete patterns.
Contents¶
| # | Document | Description |
|---|---|---|
| 1 | Harness Engineering | Evaluation vs agent harnesses, guardrails, prompt-injection containment, sandboxing, verification, budgets |
| 2 | Loop Engineering | The tool-use loop, ReAct, plan-and-execute, evaluator-optimizer, orchestrator-workers, termination, context management in long loops |
Key Insight¶
- Harness engineering answers: How do we make the agent safe, observable and controllable?
- Loop engineering answers: How do we make the agent make progress towards a goal, and stop at the right time?