AI Insights · Oct 22, 2026 · Partially verified

OpenAI, SpaceXAI and Google build the same agent factory

All three describe long-running coordinator agents that react to events and hand work to specialist agents. People still review the result, and guardrails are temporary and tracked centrally.

Claim

Three independent vendors describe the same next step: a pipeline of specialised long-lived agents triggered by events, with human review biased toward merging, and guardrails treated as temporary, centrally logged shims that get deleted when they stop firing.

Evidence

  • OpenAI software factory (Pragmatic Engineer, 2026-09-15, partly paywalled): nine-step pipeline. Codex gathers context from Git, Slack, Notion and Datadog; implements; babysits CI; routes risky PRs to a performance harness; domain-specialist review agents auto-approve low-risk changes; a rollout agent builds its own dashboard; "Perf Factory" agents triage alerts; "Sevbot" suggests but does not execute incident mitigations. Non-engineering Codex usage went from about 0 to 90% in four months; some CI and VCS systems saw about 10x load.
  • Codex AGENTS.md is layered (global, repo root, nested; closer wins), capped at 32 KiB combined, built once per run.
  • Grokbot (from talk summaries, secondary): long-lived coordinator bots orchestrating fleets of workers; prompting being added back for new domains (slide decks, travel booking, act-versus-ask before emailing colleagues); internal primitives "subscriptions" (wake on Slack post, CI failure, time of day) and "tracked shims" (centrally logged nudges; an agent files a PR to remove a shim once it stops firing). Katz's guidance: build features cheap to add and remove, invest in composable primitives like subscriptions, expect harnesses to "get messier before they get cleaner". The open xai-org/grok-build repo (Rust, Apache-2.0, 27.5k stars) is SpaceXAI's coding harness with ACP, hooks, skills and plugins; its link to Grokbot is UNVERIFIED.
  • Google Antigravity 2.0 (I/O, 2026-05-19): standalone agent-manager desktop app with parallel agents, custom subagent workflows, scheduled background tasks and artifacts; Gemini API Managed Agents bundle inference, tools and code execution. Reported replacement of Gemini CLI for free users on 2026-06-18 is secondary. UNVERIFIED.
  • Garry Tan, "Thin Harness, Fat Skills" (2026-04-11, about 1M views): the harness should only loop the model, read and write files, manage context and enforce safety; intelligence moves up into Markdown skills, execution down into deterministic tools. Criticises 40-plus tool definitions and slow MCP "God-tools" (Playwright CLI about 100 ms versus Chrome MCP about 15 s for a screenshot-find-click sequence). Text via a mirror; UNVERIFIED verbatim.

Open questions

  • Grokbot's architecture beyond summaries; the "@bot" product name.

Sources

AI-assisted research: gathered and summarised with AI from the sources above, each claim checked against them.

Related

Book a demo