Devin's Code Scans run repo-wide sweeps as agent map-reduce

Cognition's Devin now takes on repo-wide cleanup work as a single broad instruction, something like "reduce maintenance burden," instead of a list of files to touch. The feature, called Code Scans, published September 16, runs four phases. Plan writes rules for what counts as relevant code. Shard splits the matches into batches. Map fans parallel agents out across those batches. Reduce merges everything back into one report, deduplicated.

The stakes are in the numbers Cognition is citing. A Rust debug build went from 58.6 seconds to 21.0, a 64% cut. An Ahrefs health score moved from 87 to 92, with slow pages down 73%. A team at Philips reported a 96% PR merge rate on changes the scan proposed. Those are Cognition's own figures, not independently verified, but they're specific enough to hold the company to if the tool underperforms in practice.

What makes this worth copying isn't the product itself, it's the decomposition. Splitting a vague, repo-scale goal into a plan step, then shards small enough for one agent's context window, then a merge pass that dedupes overlapping findings, is a pattern any engineer can run by hand with existing tools. You don't need Devin's product to get most of the benefit: write the rules first, batch the matches, run agents per batch, then reconcile.

Cognition's post doesn't say how Code Scans handles conflicting fixes proposed by different shards, or what happens when a repo's structure doesn't shard cleanly, like a monolith with deep circular imports. Anyone testing this on their own codebase should watch for how the reduce step resolves overlap, since that's the step most likely to break at scale.