Agentic software factory: history, then check the utilities
TrueFoundry’s field guide: factories aren’t new; agent-native ones need model supply, credentials, QC, telemetry, and spend metering—not just more agents.
Source: The Agentic Software Factory, Explained: History, Architecture, and Enterprise Controls
Useful taxonomy dump with citations: Bemer → SDC → Microsoft Software Factories → BMW-scale CI, then 2026’s agent-native revival (Factory.ai, BCG Platinion, Spotify Honk, OpenAI harness post). Distinguishes coding agent (worker) vs factory (production system) vs loop engineering (one verify-retry cycle) vs platform engineering (golden paths for humans).
The editorial hook: inspect utilities first—model gateway, credential custody, guardrails, traces, budgets. Agents increase blast radius; controls must sit on mediated boundaries agents can’t bypass. Checklist questions before scaling are the durable part.
Treats vendor productivity claims as heterogeneous anecdotes, not a benchmark set.
Why it matters for code review
Enterprise review of agent output fails open if credentials, policy, and evidence aren’t in the path. “Add a review bot” without utilities is craftsman tooling dressed as a factory.