How to Generate Mergeable Code with a Context Engine

Radiologists call the failure satisfaction of search: you read a scan, find one indicator, stop looking, and miss the others that would have changed the diagnosis. Peter Werry says agents do exactly this to a codebase. Attach a wiki and an agent will search it, land on something plausible, and quit, which is why he argues access to information is not understanding. Before agents you were the context layer, trawling discussions, reading code, carrying the tribal knowledge yourself. An agent is closer to an expert engineer on their first day, rediscovering how you build, test, and deploy on every task, then forgetting it.

A million token window does not fix it: the context does not fit, and the agent gets distracted. He demos the alternative on Unblocked's own repository, where a question about an internal component returns an architecture diagram that did not exist before, sources attached so a human can check it. Then the same optimization plan twice in Claude Code, with the context engine and without. With it, under a dollar and about a minute; without, roughly double the time and more cost, because the agent has to discover things and discovers the wrong ones, so later steps run on bad assumptions and loop. The compounding is the point, not the first task. He closes on a review agent that boosts comments by reviewer seniority, a drop in flagged issues traced back to the Slack thread that explained it, and two open source pieces: a query engine over your GitHub history and a social graph showing thin review coverage.