A model alone isn't an agent: a first-principles handbook on agent harness engineering

techNmak · x · 2026-10-04

The author argues that harness engineering is the most important yet underappreciated part of building reliable AI agents, often reduced to just prompts and tool calling. A capable model by itself is not an agent — a surrounding control layer must decide what context the model sees, which tools it can use, how state survives between steps, which actions need approval, how failures are returned, when to retry or stop, and how success is verified.

This layer cuts across the whole agent stack: tool design shapes how easily a model acts, context policy determines what survives long tasks, sandboxes and permission boundaries constrain generated code, and more tools or scaffolding can actually make an agent worse by adding ambiguity. It also explains why two agents using the same model can behave very differently.

The author compiled a technical handbook working through this from first principles: agent loops, Agent-Computer Interfaces, tools, MCP, skills, active context vs durable state, compaction, sandboxes, permissions, credentials, prompt injection, execution budgets, verification, evaluator loops, long-running recovery, idempotency, subagents, observability, and agent evaluation — grounded in original research and current framework docs.

Original post →

A model alone isn't an agent: a first-principles handbook on agent harness engineering

More from coding & agent