The post’s thesis is in its title: after a year of watching coding agents ignore instructions, run dangerous commands and go in circles, the authors kept concluding that the problem was not the model but its configuration. Smarter models will remove some failure modes, it argues, but will then be given bigger problems and fail in new, unexpected ways, so the useful question is how to get the most out of today’s models. It summarizes this as “coding agent = AI model(s) + harness”, where the harness is the agent’s configuration surface — skills, MCP servers, sub-agents, memory, AGENTS.md files and the like.
It says the term Harness Engineering was coined by Viv, describes it as the practice of leveraging these configuration points to improve a coding agent’s output quality and reliability, and positions it as a subset of Context engineering — the part that uses harness configuration points to manage coding agents’ context windows. The rest of the post goes through each configuration surface in turn with HumanLayer’s own practices.
Key Points
- CLAUDE.md / AGENTS.md files (see CLAUDE.md and AGENTS.md) are the first thing to customize: markdown files at the top of a repository that the harness injects into the agent’s system prompt. The post restates its own earlier advice from Writing a good CLAUDE.md — do not auto-generate the file, keep it concise and universally applicable, use progressive disclosure — and says HumanLayer’s own CLAUDE.md is under 60 lines.
- MCP servers are for tools. Their tool descriptions are injected into the system prompt, so too many tools push the agent toward the “dumb zone”, and an untrusted server is a prompt-injection vector. Where an MCP server duplicates a CLI well represented in training data (GitHub, Docker, most databases), the authors found prompting the agent to use the CLI works better; they replaced the Linear MCP server with a small context-efficient CLI documented in their CLAUDE.md (see Model Context Protocol).
- Skills are for reusable knowledge and tools through progressive disclosure: the agent gets specific instructions or tools only when it decides it needs them (see Agent Skills and Progressive Disclosure). The post warns that skill registries have been caught distributing malicious skills and advises reading a skill before installing it.
- Sub-agents are for context control, not for role-play: “frontend engineer” or “backend engineer” sub-agents did not work for the authors, whereas using sub-agents as a “context firewall” — the parent sees only the prompt it wrote and the condensed result — keeps the parent thread coherent across many context windows, which the authors tie to models performing worse at longer context lengths (see Context rot). They also use a cheaper model for sub-agents than for the parent session to control cost.
- Hooks are for control flow: user-defined scripts run at points in the agent’s lifecycle, used for notifications, automatic approval or denial of tool calls, integrations and verification. Its example hook runs a formatter and type checks when Claude stops, stays silent on success, and returns exit code 2 on failure so the harness makes the agent fix the errors (see Agent Hooks).
- Back-pressure: the post states that the likelihood of solving a problem with a coding agent is strongly correlated with the agent’s ability to verify its own work, calls its tests and verification mechanisms one of the highest-leverage things the team has built, and insists they be context-efficient (see Backpressure and Context-Efficient Backpressure for Coding Agents).
- On post-training, it acknowledges that a model may perform better in the harness it was post-trained on, but argues this cuts both ways because models can be over-fitted to their harness, so it is not a reason to leave the harness uncustomized.
Context
The recommendations come from HumanLayer’s own experience over “dozens of projects and hundreds of agent sessions”, and the post closes with the team’s own list of what did and did not work: start simple and add configuration only when the agent actually fails, rather than designing an ideal harness upfront or installing many skills and MCP servers “just in case”. It presents itself as a companion to Viv’s own writing on harnesses, adding two levers it says he does not emphasize: hooks for deterministic control flow and skills for progressive disclosure of knowledge. Related pages include Agent Harness and Advanced Context Engineering for Coding Agents.