Researchers have built an AI framework that looks for trouble you did not ask it to find.
Most AI assistants only fix what you point them at. Ask about one bug, and they ignore the three others sitting nearby in the same file. A new paper describes TIDE, a template-guided iterative framework designed to proactively surface multiple hidden problems in a user's documents, code, or workspace, rather than just the one issue a person happened to flag. It works in rounds: each pass surfaces a small batch of candidate problems while building on what earlier rounds already found, using reusable "thought templates" distilled from past cases to keep predictions anchored to recognizable problem types instead of vague guesses. The researchers tested it on personal workspaces and software repositories, across four different model backbones, and reported it beat both single-shot prompting and parallel multi-agent baselines at finding, identifying, and resolving issues.
The interesting part is the framing, not the framework. Coding assistants and document copilots are graded almost entirely on how well they answer the prompt in front of them, which means an agent can be excellent at the one bug you mention and blind to the five nearby ones you did not. TIDE is a bet that the next competitive edge for these tools is not answering questions better but noticing what nobody asked about, the difference between a spellchecker and an editor who flags the paragraph that does not make sense.
That same instinct is exactly what makes proactive assistants risky: a tool that goes looking for problems will occasionally invent ones that are not there, and nobody wants an editor who rewrites sentences that were fine. Iterative discovery grounded in evidence is a reasonable guardrail on paper, but whether it holds up outside a benchmark is a separate question entirely.