The reframe worth stealing from Databricks’ writeup is that runaway coding-agent spend is an ops problem, not a model-selection problem. Most teams stop at “pick a cheaper model.” The bigger lever is the harness — context hygiene, caching, tool orchestration — which they tuned for a ~50% cut in generated tokens with no measured quality drop.
- 💡 The harness owns the bill. Same model, same codebase, wildly different cost depending on how the harness manages context and cache. Model choice is one line item, not the whole invoice.
- ⚡ Cache tuning is free money. Hand-tuning default cache settings to raise the hit rate cut token cost hard — the least glamorous optimization with the best return.
- 🎯 Route, don’t standardize. Their Unity AI Gateway smart router shaved 30%+ off average task cost by sending easy requests to cheaper models and keeping quality flat.
- ⚠️ Downshift instead of capping. When a developer crosses a spend threshold, drop them to a cheaper model rather than blocking them — progressive friction beats a hard wall, the same way graceful degradation beats a hard failure under a latency budget.
- 🔍 Guard against harness lock-in. Harnesses are increasingly co-designed for one model family; a meta-harness (they open-sourced Omnigent) keeps you able to swap models without rewriting your workflow.
None of this is exotic — it’s the same routing, fallback, and budget-guard discipline that agentic systems already need in production. The HN discussion is worth a skim for teams weighing the same tradeoffs.
Here’s the part I’d push on: “no quality degradation” is only as trustworthy as the eval harness measuring it. If you can’t prove a leaner harness or a cheaper model didn’t quietly regress your outputs, you haven’t cut costs — you’ve deferred the bill to a worse code review. Can your eval catch a 5% quality slide before your developers do?