Most teams point agents at small, well-scoped work. A ticket goes in, a tidy PR comes out, a human reviews it, it merges. Our largest projects don't work that way. On Cosmos, once a human has reviewed and approved the design doc, we hand the entire design to a single worker agent and let it build the whole thing as one PR. It implements the design, fixes its own CI failures, and works through review comments until the change is ready to merge. No ticket-splitting, no relay between agents. The PRs are large, often thousands of lines, and we prefer them that way. One agent holding the full approved design writes a more coherent change than several agents each working from a slice of the context. The seams you get from splitting, the mismatched assumptions between unit three and unit seven, don't show up when one agent carries the whole picture.

A few we shipped to production like this, all on systems with thousands of active users, none of which caused an incident.

The human reviews the design before any code exists. Our Code Review experts then read the full diff in minutes and check it against that approved design. Implementation size stopped being a constraint the moment a person was no longer the one reading every line. The unit of work is no longer the PR. It's the project, and one agent can now carry a project from approved design to merged code on its own. Full write-up: augmentcode.com/blog/accelerat…