98.4% of a Production Agent Is Harness, Not AI. The Model Barely Matters.

VILA-Lab published a systematic teardown of Claude Code's TypeScript source on arXiv (2604.14228), and the headline number stopped me cold: 98.4% of the codebase is deterministic operational infrastructure. The AI decision logic is 1.6% of the system.

I've been building my own autonomous pipeline for months now, and this tracks with everything I've experienced. The interesting engineering isn't the model call. It's everything surrounding it. The paper maps 13 design principles derived from 5 human values (safety, efficiency, quality, cost, developer experience) down to specific implementation choices, and the architecture patterns are immediately useful.

The most actionable finding: before every single model call, five compaction shapers run in sequence, cheapest first. Budget Reduction strips what it can for free. Then Snip removes tool results that exceeded their useful life. Microcompact compresses recent content. Context Collapse merges older conversation turns. Auto-Compact rewrites everything remaining into a tighter summary. Each one runs only if the previous didn't free enough space.

This is a waterfall, not a single compression step. And it's deterministic. No model involved.

The implication for builders is direct. If you're spending your time evaluating which model to use, you're optimizing the wrong 1.6%. The harness decides what context reaches the model, when tools get called, how errors recover, and what gets remembered between sessions. As frontier models converge in reasoning capability (and they are converging, fast), the harness becomes your only real competitive advantage.

I see teams pouring weeks into prompt engineering while their context management is a single system message that grows unbounded until the window fills up. That's like tuning a race car engine while driving on flat tires.

If you're building agents today, read this paper and ask yourself: do I have a compaction pipeline? Do I know what percentage of my context window is signal versus accumulated noise by turn 50? The AICC enterprise report published this week notes that context window compounding can balloon a session from 5K to 200K tokens by turn 50. That's not a model problem. That's a harness problem.

Build the harness. The model is a commodity.