Invest in your harness

An agent harness is the system around the AI model that helps it do real work on your project. We think about ours in eight parts: context (what the agent knows about your code and conventions), provenance (typed links that record why each change exists), capability (tools that act on and observe live state), workflow (slash commands, plan-then-execute, subagents, skills), restraint (rules, permissions, allowlists), verification (tests, type checks, fail-first reproductions, AI tool simulators), a visual interface (the workspace where work happens and gets reviewed), and coordination (how the human keeps track of many agents working in parallel). The model is interchangeable. The harness is where your durable investment lives.

Claude Code and Codex are themselves harnesses. They wrap a frontier model with a system prompt, a tool set, a permission system, and an execution loop. Your team provides a second harness on top of that: the workspace, the linked context, the workflow, the rules, the verification loop, the multi-agent coordination, and the tools that are specific to your project.

Frontier models flip the leaderboard every few weeks. Recent studies from Stanford and Tsinghua show that the orchestration code around the model drives more performance variation than the model itself: the same model can produce a six-times gap in result quality depending on the harness it runs in. Investment in your harness compounds and survives model churn. Investment in tuning prompts for last quarter's model does not.

Start with a CLAUDE.md or AGENTS.md at the root of your project that captures your real conventions and hard rules. Add path-scoped rule files for areas with special concerns. Wire up at least one tool that lets the agent verify its own work, like a test loop or a screenshot tool. Adopt a workspace like Nimbalyst that gives you a linked provenance graph, workflow scaffolding, visual editors, and multi-agent coordination out of the box, so the agent and the human can work from the same artifacts.

The provenance graph records persistent, typed links between the artifacts that matter. Tracker item to plan, plan to spec, spec to diagram, diagram to session, session to diff, diff to files, decision to the work that forced it. Without it, the connections between work live only in human heads and an agent cannot traverse them. With it, both human and agent can pick up where the last session left off in a single traversal.

Not much, at first. Coordination is the pillar that compounds as you start running more than one agent in parallel. A kanban of sessions, workstreams that group related work, and a meta-agent that can spawn siblings or supervise long-running background tasks let the human keep a single overview across many agents instead of opening more chat windows.

No. Many of the pieces of a good harness, like CLAUDE.md, path-scoped rules, and tool definitions, can be built up inside any project. Nimbalyst is one open-source example of a workspace that already includes the provenance graph, workflow, verification, visual interface, and multi-agent coordination parts. Adopt it whole, copy ideas from it, or use it as a reference while building your own.