Harness Engineering

Runtime + Delivery · 7 Chapters

The machinery that turns model capability into controlled work — the runtime harness beneath a model owns its loop, context, compaction, tools, and system prompt; the delivery harness above one or more runtimes owns work, policy, evidence, and authorization. Not vendor features (Toolkit). Not portable working techniques (Patterns). Not the trust and privacy surface (Security).

The 7 Chapters

Read top to bottom. The sequence starts inside one runtime and ends at the delivery layer above one or more runtimes.

The Agent Loop

The core run loop: prompt, model, tool calls, results, repeat — and the stop conditions that end it.

Read

Context Window Economics

The token budget as the scarce resource: what occupies the window, the cost model, and what to evict.

Read

Compaction & Continuity

What survives a context-overflow boundary: summarization fidelity, the handoff, and designing for it.

Read

Tool Result Curation

The signal-to-noise problem: verbose tool output, truncation, deferral, and protecting the window.

Read

System Prompt Architecture

Where steering actually lives: instruction hierarchy, the harness preamble, and layered precedence.

Read

Building a Harness

Rolling your own loop with the Claude Agent SDK — when it pays off, and when to use the managed one.

Read

Delivery Control Above the Agent Loop

The layer above runtime harnesses: exact-candidate state, independent evidence, bounded authority, and accountable delivery decisions.

Read

Where This Connects

Harness Engineering spans runtime execution and delivery control. These patterns, toolkit features, and security controls meet it at different authority boundaries.