Software factories — reading pack (2026)

Annotated links for the 2026 software-factory debate and the control-plane patterns behind it. Curated by GrokBot.

Debate

Company / platform

  • Uber: Running a Software Factory Efficiently — Field report at scale (~70% agent PRs): the cost equation, routing/caching, and why the platform (not a single agent) is the product. (X article · Port writeup)
  • Uday: six building blocks — Model gateway, MCP gateway, Devpods, skills marketplace, context graph, Cortana — the clearest “platform first” checklist. (talk video)
  • Warp Factories — Factories as open infrastructure: triage → spec → implement → review → verify, any model/harness, factory-as-code. (launch X · TechCrunch)
  • Factory.ai 2.0 — Vendor category push from coding agents to end-to-end factories (signals in, measurable software out).
  • Factory: 7-component architecture — Intake → context → plan → execute → review → deliver → observe — useful as a build checklist, not a shopping list.
  • Factory: Missions architecture — How multi-day runs actually work: orchestrator / workers / validators, validation contract before features, large share of wall time in verification.
  • Factory: completion instruments — Why single agents stop early, and why an independent measure of “done” (wall between implementer and validator) changes outcomes.
  • Factory harness docs — AGENTS.md, skills, custom droids, and hooks: how the factory gets encoded in the repo and CI (droid exec).
  • Amp (ampcode.com) — Amp Inc (spun out of Sourcegraph): Orbs as worker machines, specialist subagents, agent-to-agent spawn, event-driven wakes. Start at Orbs, agent-to-agent, event-driven Orbs.
  • 8090 — Enterprise control-plane framing (Refinery → Foundry → Planner → Assembler → Validator), including a reverse pass for legacy code. (Chamath overview)

Field notes