The Harness Is the New IDE

#ai#agents#harness#tooling#kodr

Something I keep coming back to: the harness is the new IDE.

For twenty-odd years the IDE was where the leverage was. Better autocomplete, better refactoring, better debugging - pick your editor, learn it deeply, get faster. That equation is breaking down. I still have an editor open, but I look at it less every month. The thing I actually configure, tune, and think about is the harness.

By harness I mean the loop around the model: Claude Code, Codex CLI, Antigravity CLI, and the rest. The thing that manages context, runs tools, applies permissions, decides when to plan and when to execute. The model is the engine, but the harness is the drivetrain - and right now, harness design is where the interesting differences are.

Models matter too, obviously. A better model is a force multiplier on everything. But you don’t get to design the model. You do get to choose and shape the harness. Your AGENTS.md file, your hooks, your permission rules, your sub-agent setup, how you split work, when you compact - that’s the part of the system you actually own. Two people on the same model with different harness habits get wildly different results. I see this constantly.

I learned this properly by building one. Several, actually (a rust harness I never cleaned up for release, kodr, then kodr2): a tiny harness with “zero node dependencies”, specifically designed to run against local models. Writing even a toy version forces you to confront everything the good harnesses paper over. Context budgets. Tool schemas. What “permission” even means when an agent can run shell commands.

Meanwhile the (for me) IDE is quietly becoming a viewer. I use it to read diffs the agent produced, not to type code - mostly tweaks. The features I used to care about - autocomplete, snippets, even most refactoring tools - are solving a problem I no longer have. My results are more about the outcome than how I get there. What I want from an editor now is a great diff view and a terminal!

The skill that’s appreciating isn’t editor mastery, it’s harness mastery: writing the guidance files, working with specs, designing the feedback loops, knowing when to hand the agent a plan versus a goal. That’s the thing worth getting good at.

Not a prediction, just an observation of where my own hours already go.

Links: