The plumbing that connects an AI to real actions is harder to edit than the AI itself — one behavior hides in a dozen places.

This explores why the harness — the infrastructure layer that wires a model's outputs to tools, prompts, and staged actions — ends up more tangled than other agent code. The starting point is that agent code isn't one thing. It splits into three layers with different failure modes: the model's trained internal capability, the system-provided harness that turns outputs into actions, and the artifacts the agent writes while running What are the three distinct layers of agent code?. The harness is the awkward middle layer, and its tangle comes from a specific structural fact about how it stores behavior.

The core reason: a single *behavior* in a harness is smeared across many *locations* in the code. Prompts, tool definitions, and pipeline stages each hold a piece of any given behavior, so there's a mismatch between how you'd describe a change ('make it stop retrying dead endpoints') and where that behavior actually lives structurally. The hardest part of changing a harness isn't writing the edit — it's finding every place the behavior is implemented. That localization problem, not code generation, is the real bottleneck Why is finding distributed behavior code so hard?. Tangle here is literally distribution: no one file is the behavior.

That framing pays off when you try to fix it. If you reorganize the repository around runtime behavior — an explicit map from behaviors to the code that implements them — even a weaker planner can locate the right spots as well as a much stronger model, with double-digit win-rate gains and fewer tokens spent hunting Can explicit behavior maps help weaker planners compete with stronger models?. In other words, the tangle isn't intrinsic to the task; it's an artifact of organizing code structurally instead of behaviorally.

There's a second kind of tangle worth naming: when the harness tries to improve itself, its 'learning' is easy to overstate. Inspecting evolved harness edits shows most of them just cache task-specific fixes the agent could have rediscovered in a single run, rather than distilling reusable strategy Do harness edits learn reusable strategies or memorize task fixes?. And because harness evolution is itself a search loop, apparent design gains are confounded with raw search effort unless you compare against an equal-budget baseline on held-out tasks How should we measure gains from automatic harness evolution?. So the harness is tangled both to read (behavior is distributed) and to credit (improvement blurs with memorization and search).

The deeper lesson the corpus keeps circling: harness tangle tracks how integration is done. Production teams find that protocol-mediated tool access (like MCP) introduces non-deterministic failures through ambiguous tool selection, and that replacing it with explicit direct function calls restores predictability Why do protocol-based tool integrations fail in production workflows? — with 85% of teams building custom agents rather than adopting frameworks Should coordination protocols wrap existing systems or replace them?. The through-line: harnesses get tangled when a behavior has no single home. Whether the fix is behavior-centric repo organization or deterministic direct calls, the win comes from collapsing 'where does this behavior live?' back down to one answerable place.