The model eats everything. A harness should get smaller as the model gets better. The test is to ask what the model would still have to guess if the instruction disappeared. A model can figure out how to edit a file. It needs evidence to know which settings the repo overrides, where its issues belong, and which version of the diff someone reviewed.

Looking at OpenCode and the Harness Leaderboard gave me a way into that question. A leaderboard compares systems as they were run. To know whether a particular prompt, tool, or context policy helped, I’d have to vary it while holding the rest steady. This round changed two parts of Kit that I could inspect directly [1].

First, Kit gives global and repo settings the same shape. Global preferences use the standard XDG config location and can still be read from older paths. The repo’s .kit.yaml wins over global defaults, and issue routing belongs to the repo. A reconcile plan could carry personal defaults into the repo file, so Kit now keeps effective reads separate from repository writes. I also made kit guide a short index, with flags and examples available through kit guide <command>. It still draws from the CLI’s help [2].

Second, every PR still starts with a fresh independent review. The PR now records the commit or diff, the scope, and the outcome. When later edits change behavior, permissions, acceptance criteria, or what a test considers a pass, an independent reviewer reads the affected delta. A typo gets a lead diff check. The unit of review is the change that could invalidate a conclusion [3].

This round supplied its own example. A follow-up reviewer found that a valid YAML key with a binary tag could decode to issues after Kit had already filtered the global keys. In a repo with no issue settings, the wrong tracker could be selected. Kit now normalizes the keys before filtering and tests that loading and scaffold generation agree [4].

That is one caught bug. The cost of asking for a focused second opinion on every substantive delta is still an open question. I’d run the same coding tasks several times with the same model under current Kit, a trimmed Kit, and a bare agent. Score correctness and delivery first, then count time, tool calls, cost, and human interventions. That comparison can tell me whether the trimmed harness earns its place. To judge any one rule, I’d repeat the comparison with that rule as the only change.

This is the first Coding-Agent Harness Research note. The series will pair changes to Kit with the question that prompted them and the evidence available when they shipped.


[1] OpenCode source and the Harness Leaderboard
[2] Kit #304, config provenance and agent guide
[3] Kit #306, follow-up review coverage
[4] Tagged YAML key fix and regression