"I fixed it, take a look."

Over the past year, every developer using AI to write code has heard this line one too many times. The Agent runs for two minutes, confidently submits results, you pull the code -- it doesn't compile, tests are all red, edge cases completely ignored. "Can write" doesn't mean "can trust" -- that's the core bottleneck preventing AI coding from achieving true lights-out, production-scale deployment.

On July 1, the open-source AI Agent project OpenSquilla shipped version 0.4.0 with a direct answer: AI no longer stops at the verbal delivery of "I fixed it." Before handing back results, it first runs tests to produce a verifiable evidence chain for itself.

Red-Green-Regression: Three Gates, Fail Any and Auto-Retry

OpenSquilla's "self-verification" mechanism is an independent red-green-regression evidence chain, executed in three steps.

Gate one: Red test characterization. The AI first writes a test guaranteed to fail, precisely reproducing the bug to fix or the feature to add. Only if this test fails does it prove the AI actually understands the problem, rather than making random changes. Gate two: Green test verification. The AI implements the feature or fixes the bug, turning the test from red to green, proving the problem is genuinely solved. Gate three: Regression safety net. Run the project's existing test suite to confirm the new changes didn't break anything else.

All three gates must pass for delivery; fail any and it gets sent back for automatic revision until it passes. Supporting mechanisms include isolated execution -- all changes happen in an isolated copy, only merged back to source after passing acceptance, preventing Agents from accidentally deleting your production code.

The official demo used Andrej Karpathy's micrograd project: the AI needed to add "correct gradient computation" to this minimalist autodiff library -- and miscalculated gradients don't throw errors or crash; they silently make the model learn worse, the kind of bug hardest to spot by eye. After OpenSquilla completed the red -> green -> regression sequence, the team compared the results side-by-side with PyTorch on the same problem and found forward values and every single gradient matched to 10 decimal places. Not "the AI says it's right," but "it matches the authoritative answer exactly."

Costs Drop 9x, the Agent Harness Race Heats Up

Beyond self-verification, OpenSquilla's other selling point is cost efficiency. Against a backdrop where mainstream Agent frameworks are driving up model calls and token costs keep climbing, it uses local intelligent routing -- automatically selecting models by task complexity, loading skills on demand, retrieving memory on demand, preprocessing tool results -- to crush costs before the call is even made.

Official data shows its intelligent routing beats the general-purpose gateway OpenRouter by approximately 4.4 percentage points in routing accuracy while costing roughly 75% less; quality is on par with flagship models on comparable tasks while costing roughly 9x less. Internal testing shows 60-80% comprehensive cost reduction on routine scenarios. Within weeks of launch, GitHub stars surged to the thousands, and the company closed its first funding round just months after founding.

OpenSquilla isn't alone. Over the past three months, Agent Harness -- the infrastructure layer that provides execution environments, verification mechanisms, and toolchains for Agents -- has become the new hotspot in AI coding. From Cursor's Agent mode to Claude Code's CLI tool, from Devin's fully automated DevOps to various open-source Agent frameworks, everyone is solving the same problem: how to let Agents reliably complete real-world engineering tasks without a human babysitter.

Senior SWE-Bench, which hit the Hacker News front page the same day, points in the same direction -- the new benchmark doesn't test Agents on LeetCode problems; it evaluates their performance as "senior engineers" in real codebases. This means the industry's judging standard has shifted from "generation quality" to "delivery reliability."

As the AI coding tool wars rage into the second half of 2026, simply competing on model capability is no longer enough. Whoever builds a reliable verification loop, whoever drives hallucinations and error rates down to production-grade levels, whoever crushes costs while maintaining quality -- that's the tool that actually helps developers get work done, rather than an "intern" you have to constantly watch and clean up after.