The Steering Loop: Harness Engineering with OpenCode for the Agent-First Era

The bottleneck in AI-assisted software delivery is not the model. It is the environment around it.

In early 2026, OpenAI's Harness team shipped a million-line production codebase with three engineers and zero lines written by hand. Their diagnosis was blunt: when agents failed, the problem was rarely capability. It was missing structure, feedback, and mechanical constraint. Every gap they found was encoded back into the repository. The environment improved. The agents improved with it.

This compounding cycle is harness engineering, and in this session we demonstrate how to implement it using OpenCode, the open-source AI coding agent trusted by millions of developers and designed to work across models, providers, and deployment environments.

Harness engineering has two core components. Guides are feedforward controls that steer the agent before it acts—LSP integrations, structured AGENTS.md files, architectural guardrails, and reusable Skills. Skills allow teams to encode institutional knowledge directly into the agent's workflow: how to write tests, review pull requests, bootstrap services, enforce team conventions, and apply organization-specific engineering practices. They are versioned, reusable engineering assets that improve over time.

Using OpenCode as the implementation platform, this talk provides a practical blueprint for building an agent-ready engineering environment. Every pattern demonstrated is model-agnostic and can be adopted regardless of which LLM provider your organization uses.

Attendees will leave with a clear mental model for harness engineering, a forkable OpenCode-based harness architecture with reusable Skills, and a pragmatic understanding of where human judgment remains essential in the agent-first era.