Microsoft Research: LLMs Silently Corrupt 25% of Your Documents in Long Workflows

The same week we're celebrating AI rewriting a million lines of code, Microsoft Research dropped DELEGATE-52, and it's the cold shower this industry needs.

The benchmark simulates long delegated workflows across 52 professional domains, from coding to crystallography to music notation. The finding: even frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4) corrupt an average of 25% of document content by the end of long workflows. The errors are sparse. They compound silently. And agentic tool use doesn't help.

451 HN points. Not because it's surprising, but because every practitioner has felt this and now has numbers.

The word "silently" is doing heavy lifting here. These aren't errors that throw exceptions or fail tests. They're the kind of corruption where a number changes, a clause disappears, a constraint gets softened. The model doesn't flag it. You don't notice until something downstream breaks. If you're lucky enough to notice at all.

Three variables make it worse: document size, interaction length, and distractor files. Bigger documents, longer chains, more noise. This is the exact trajectory every "autonomous agent" architecture is optimizing for. More context, longer runs, more files in scope. The DELEGATE-52 results say that trajectory leads directly into silent data corruption.

Here's where I think builders need to change behavior today. If you're running any multi-step agent workflow that modifies documents, you need checkpoint verification. Not at the end. At every step. Diff the output against the input and verify that only intended changes were made. Yes, this is expensive. Yes, it's slower. And yes, it's the only thing that catches 25% corruption rates before they compound into something you can't recover from.

The uncomfortable pairing with the Bun story is this: AI can rewrite 960K lines with 99.8% accuracy WHEN YOU HAVE TESTS. Without that verification layer, you get the DELEGATE-52 numbers. The test suite isn't optional infrastructure. It's the difference between a tool and a liability.