The hosts explain harness engineering: making coding agents run autonomously for 40–50 minutes, writing, testing, and proving code production-ready, by giving them the right environment, signals, and enforced guardrails. OpenAI’s case: a repo started in August 2025 by three engineers grew past a million lines in months, with productivity rising as the team grew, because instead of telling the agent to “try harder” they asked what capability it was missing and built it into the enforced flow. Practical recipe: product docs as files in the repo, skills that say when to update themselves, git hooks enforcing 80% test coverage, TDD so agents can’t cheat, agent-launchable isolated environments, and alignment in three steps (spec → plan → proof of work, e.g. a Playwright video of the feature working). Also: tracing that agents can query, why custom org-specific skills beat popular generic ones, and cost — providers moved to token billing, but don’t blindly minimize tokens; measure with hooks/OpenTelemetry. Closing take: harness gains compound; measure productivity per token over time.
Show: langtalks
Concepts: harness-engineering · agent-ready-codebase · test-driven-development · ai-verification · ai-guardrails · context-engineering