Reliability Is the Product of the Harness, Not the Model

repoaiagentsharness-engineeringagent-loopsverificationeducationjoelclaw

Maps directly to joelclaw's agent-loop/workload-rig pattern: instructions, scoped state, verification gates, and lifecycle control around coding agents.

Walking LabsLearn Harness Engineering is a project-based course about making AI coding agents reliable by building the environment around them: instructions, state, verification, and lifecycle control.

The useful frame is simple: the model isn’t the whole system. The repo points at OpenAI’s harness engineering essay and Anthropic’s long-running agent harness work, then turns that into 12 lectures, 6 projects, multilingual docs, PDFs, and a harness-creator skill that scaffolds files like AGENTS.md, feature lists, init.sh, and verification workflows.

That’s the clever bit. It treats “agent got confused” as an engineering failure, not a prompt vibes failure. Context loss, premature “done,” missing tests, and sloppy cleanup become things the harness has to prevent.

For joelclaw, the interesting move is to compare its course structure against our own agent-loop setup: repo instructions, skill loading, workflow state, reviewer/judge gates, and proof that work actually passed. If a fresh agent can’t start, continue, verify, and stop cleanly, the harness is still leaking.

Key Ideas