Your Agent Harness Just Became a Commodity. Good.
OpenAI shipped the Codex harness as an API and charged nothing for it — here's why the orchestration layer your team built in 2026 was never the moat.
The Agent Loop Is Now Plumbing. Stop Paying Engineers to Rebuild It.
OpenAI just took the six months your platform team spent on context compaction, subagent fan-out, and session recovery, wrapped it in four API primitives, and priced it at zero. At Kuaray, we've reviewed enough agent architectures this year to say the quiet part: most of that orchestration code was never a differentiator. It was tax. The Agents API, in public beta since September 10th, is the invoice finally coming due.
What Actually Shipped
The managed Codex harness behind one call. Four concepts — Agent, Environment, Session, Events — and OpenAI runs the miserable parts:
| You used to own | Who owns it now |
|---|---|
| Summarization when context fills | Harness, per root agent and per subagent |
| Subagent spawn + result merge | multi_agent: { max_concurrent_subagents } |
| Crash recovery across multi-day runs | Durable sessions |
| Loading 200 tool definitions every turn | Tool search, lazy-loaded |
No extra fee. Tokens, tools, container time. That's it.
And read the customer numbers with the skepticism they deserve — they're vendor-supplied — but one is instructive. Hypha reported 86% fewer failed agent responses after separating the harness from the sandbox. Not a better model. Better plumbing. Which tells you where most of your agent failures have been living.
The Fine Print That Kills Half the Deals
Public beta means US-only data residency and no Zero Data Retention. If you're a German bank or anyone inside an EU AI Act high-risk classification, the decision just got made for you. You're not migrating.
Which is why the other announcement that week matters more than anyone noticed. On September 15th, Coder shipped Claude Code into Agent Relay — the same harness, self-hosted, network-governed, running on the customer's own infrastructure. Two vendors, same week, opposite bets on where the agent should execute.
That's not a feature race. That's the actual architectural fork of 2027: rent the loop, or run the loop where your data already lives.
What We Tell Clients
- Delete your compaction logic. Seriously. If you wrote a custom summarizer this year, it is now unpaid maintenance burden competing against a team that watches Codex telemetry all day. You will lose that race.
- Keep prompts, tools, and state models portable. Rent the harness, never the domain. An MCP server you own runs against any harness; a
sessions.createcall buried in your business logic does not. - Decide environment before model. OpenAI-hosted, self-hosted
codex exec-server, or a partner sandbox — that choice determines your compliance posture, egress story, and blast radius. The model is a config line. The sandbox is architecture. - Re-run your build/buy math on anything agentic older than six months. The baseline moved twice this quarter. Whatever you scoped in March is now more expensive than the free version.
Here's the uncomfortable truth for engineering leaders who staffed a platform team around agent orchestration: the loop was always going to be commoditized. The moat is your tools, your data boundaries, and your evals. It was never the while statement.
Schedule a Technical Architecture Review with our Strategists — before you renew another quarter of orchestration headcount.