← Field Notes

Field note · Agent engineering · March 2026

The harness around the model is the product

Model capability sets the ceiling. Harness design determines whether the system can plan, continue, verify, recover and produce consistently useful work over long horizons.

AudienceCTOs · founders · engineering leaders
PerspectiveProduction architecture and delivery
FormatOriginal analysis grounded in primary sources

Why naive autonomy stalls

A single agent with a large prompt eventually loses task structure, repeats work, damages previously correct output or declares success without verification. Longer context delays the problem; it does not solve it.

Specialise the control loop

A planner decomposes work and preserves intent. A generator implements bounded units. An evaluator applies explicit criteria and deterministic tests. Structured artifacts carry state across sessions and models.

Engineer the environment

Give the agent isolated workspaces, fast feedback, realistic test data, searchable documentation and tools with clear contracts. Add checkpoints, budgets, cancellation and recovery before extending autonomy.

How this relates to leadership

The harness encodes the same practices a technical lead uses: break down work, establish quality gates, clarify ownership, review evidence and prevent one failure from contaminating the whole delivery stream.

Executive takeaway

Buying a stronger model can improve results, but durable advantage comes from the organisation-specific harness: workflows, skills, evaluation criteria, data access and feedback loops that competitors cannot copy overnight.

Primary reading

  1. Anthropic : Harness design for long-running application development
  2. Anthropic : Effective context engineering for AI agents
  3. Anthropic : Agent Skills