Developer Tools

AI Now Writes Your Software — New Research Shows How to Trust It

Cheaper safety checks for AI-written code could mean faster fixes and fewer bugs.

Deep Dive

AI coding agents — software that writes and edits code for you — are increasingly being let loose on real, working programs. Every time one of them finishes a job, it leaves behind evidence that it did the work correctly: automated tests, type checks, math proofs, and logs of what it did. The problem is what happens next time. When the AI comes back to that same code for a new task, should it reload all of that old evidence (slow, expensive, wasteful), or skip some of it (fast, but risky if the skipped piece was the one thing holding the new change together)?

This paper offers a middle path. It treats that pile of evidence like a toolbox and asks a simple question: what is the smallest set of tools needed for this specific job? The author calls the answer an "assurance envelope." The clever part is how it checks itself — rather than trusting its own math to pick the right tools, it walks the chain of reasoning to confirm the chosen evidence actually proves the property being protected. It's a safety check on the safety check.

So what did the tests show? The author ran the method on small code examples from Rust and two AI projects, plus a synthetic test of 249 cases. For graphs with 500 pieces of evidence, the median solve time stayed under 20 milliseconds — faster than a blink. But when targets had many alternative ways of being proven, the solver choked on far smaller graphs. In other words, complexity, not sheer size, is the real obstacle. Just as useful: sometimes no valid set exists, meaning the available evidence simply cannot justify the change.

What this does not do matters too. It doesn't yet figure out which properties a change must preserve in the first place, and it hasn't shown that AI agents actually get better when given these envelopes. This is early, academic, single-author work. But it points at the bottleneck ahead: as AI writes more of our software, the cost of trusting it becomes the real constraint. Make verification cheap and automatic, and safer software ships faster.

Key Points
  • AI coding tools leave behind proof they did the job right — tests, checks, and logs. This paper asks which pieces you truly need to reload for the next task.
  • In tests, the method found the cheapest needed set in under 20 milliseconds, even with 500 pieces of evidence — but it struggled when many competing explanations existed.
  • Sometimes no combination of existing evidence proves a change is safe — a useful stop sign before buggy code reaches customers.

Why It Matters

Could make AI-written software updates faster and less buggy — meaning better apps for you and fewer crashes.

📬 Get the top 10 AI stories daily