The last mile: agents make 80% fast; finishing still does not scale

Kent Alstad’s Sep 27 opener for a KandiForge series names the asymmetry continuous factories keep rediscovering: agents make the first ~80% of building software fast — working code, passing tests, a system that looks almost done — then the last 10–20% (edge cases, integration seams, failure modes, demo-vs-production) stays human-shaped. When hundreds of agents produce work at a scale no person can match, the scarce resource is not building. It is completing.

The receipt is operational, not vibey. At KandiForge, agents had produced millions of lines across 20 repos and still needed to ship. The blocker was not missing requirements knowledge or architecture understanding — it was working every final detail across a codebase that size. Adding more reviewers and testers (agents and people) did not fix the geometry when tests ran after code was already integrated: the change gates were in the wrong place, so catches arrived as review debt rather than worktree failures. Alstad is explicit that nobody has fully solved this yet; the post is a problem statement ahead of deeper factory notes and a December BuildStuff talk.

Commenters converge on the same fix direction continuous factories keep needing: move the gate before the change leaves the agent worktree (reproducing test + minimal diff in isolation; terminate the run if the sandbox check fails), and do not let post-merge or post-integrate verification be the first hard signal.

Why it matters for code review

Always-on agent fleets without a pre-integrate complete/harden gate manufacture unfinished inventory. Review this as factory design: verification that only fires after merge is theater under volume. Steady verified momentum beats peak-session heroics only when “done” is enforceable inside the worktree — otherwise FOMO for overnight agents just widens the morning last-mile pile.