Agent Readiness Is the Real Software Factory Signal

The pitch. Davis Palmie, a member of technical staff at Factory, wants teams to move from coding agents to a "software factory" that takes bug reports and user feedback as input and ships production code. He argues coding was never the bottleneck: PR review, bug reproduction, docs, and test maintenance consume more time than writing code. This is a vendor talk, so the grand claims, like moving organizational pain from O(N) to O(1), are positioning rather than measured results.

The useful signal. The agent readiness checklist is the part worth taking. It covers validation guardrails, documented build and CI commands, fast tests, reproducible dev environments, modular code, observability, and security scanning. Agents inherit the mess teams tolerate. Stale docs and brittle tests will not stop an agent; they just make output worse and harder to verify. Any team piloting AI coding can use this list before buying more tooling.

The metric advice. Palmie argues against lines of code and token counts as success metrics and proposes signal-to-production time, human intervention count, median time to repair, shelf life of code, and cost per PR. This is the most actionable part of the talk. Token leaderboards encourage engineers to spend more, not deliver more, which is Goodhart's law in action. Engineering leaders should put outcome metrics in place before agents scale, because whatever you measure will shape what the system optimizes.

The weak evidence. The cost claims need scrutiny. Palmie says code review with GPT-5.2 can match the latest Opus models at half the price, and open-source models can cut costs 10 to 30 times. No benchmark data or production numbers support those claims in this talk. Multi-model routing and sovereign deployment are reasonable directions, but the operational cost of managing several providers, plus the integration work, is left out. Treat the numbers as directional vendor estimates, not business cases.

The governance trade-off. Palmie recommends starting narrow, for example incident triage from Sentry alerts, with humans reviewing agent diagnoses before any autonomous loop. He also says engineers must stay as validation gates for architecture and strategy. That is a sensible maturity path, but it reveals the core tension: more human gates mean less autonomy, fewer gates mean more silent drift. The right position depends on the team's risk tolerance, not on the platform's promise.

What to watch. The durable signal is that documentation, tests, and observability are becoming AI infrastructure. The software factory vision itself is a multi-year organizational shift, not a switch to flip. Teams should run a readiness assessment, pick one narrow workflow, set outcome metrics, and scale only after trust builds.