Agent Harness Design
An agent harness is the control system around a model: prompt assembly, tools, state, permissions, feedback sensors, edit formats, and failure recovery. The corpus repeatedly argues that these choices can materially change usefulness and cost even when the underlying model is held fixed.
Synthesis
- Codex frames the basic loop as repeated model inference plus tool execution, with permissions, environment context, cacheable prompt prefixes, and compaction shaping every turn. [src]
- Pi and Claude Code examples favor small cores, progressively disclosed context, structured tools, task delegation, and extensions or skills over one monolithic static prompt. [src] [src]
- Harness-engineering and Stencil accounts identify deterministic checks, useful errors, state-aware edit interfaces, and feedback loops as design variables—not incidental plumbing. [src] [src]
Coverage expansion
The Sequoia Ascent summary frames "Software 3.0" as programming LLMs through prompts, context, tools, memory, instructions, sensors, actuators, APIs, schemas, and guardrails. It is explicitly AI-generated from a video transcript and posts, so it is useful here as a stated framing rather than a primary technical specification. [src]
That framing strengthens the case for a harness as an engineered interface, not a bag of prompts. The memory-specific implementation pattern is represented by Hermes Agent, and the executable-document form by Spec Driven Development.
Practical implication
Treat the harness as an empirical product surface. Observe model behavior, make feedback fast and concrete, and keep high-risk actions behind explicit controls rather than assuming a stronger model removes system-design work.