OpenAI Will Now Run Your Agent Loop For You
OpenAI put the Agents API into public beta on September 10, and the pitch is short: build and run cloud agents with the Codex harness, fully managed. OpenAI handles orchestration, long-running sessions, and context management. You handle the part that is actually yours.
That last sentence is the whole strategic move. For two years the interesting work in agent land has been harness work, which is to say loop design, context compaction, tool routing, retry policy, session persistence. Dozens of companies exist because that layer was hard and nobody owned it. OpenAI just announced they own it, and priced it at zero. No additional fees. You pay for tokens and tools, same as before.
Shipping alongside it are hosted sandboxes, where OpenAI provisions the execution environment so agents can run code, touch files, and produce artifacts. You pick CPU, GPU, memory, storage, and secret management. Notably they did not close the door: you can point at your own sandbox or at Blaxel, E2B, Modal, DigitalOcean, or Vercel instead. Containers bill separately at standard container rates.
Docs and signup at https://openai.com/index/introducing-the-agents-api/
The HN thread was tiny, 34 points, which tells you this landed as infrastructure rather than news. That is exactly how commoditization looks from the inside. If your product is the loop, the loop is now a free managed service from the company that also sells you the tokens. If your product is what the agent knows, what it is allowed to touch, and who is accountable when it goes wrong, today was a good day.
← Back to all articles
That last sentence is the whole strategic move. For two years the interesting work in agent land has been harness work, which is to say loop design, context compaction, tool routing, retry policy, session persistence. Dozens of companies exist because that layer was hard and nobody owned it. OpenAI just announced they own it, and priced it at zero. No additional fees. You pay for tokens and tools, same as before.
Shipping alongside it are hosted sandboxes, where OpenAI provisions the execution environment so agents can run code, touch files, and produce artifacts. You pick CPU, GPU, memory, storage, and secret management. Notably they did not close the door: you can point at your own sandbox or at Blaxel, E2B, Modal, DigitalOcean, or Vercel instead. Containers bill separately at standard container rates.
Docs and signup at https://openai.com/index/introducing-the-agents-api/
The HN thread was tiny, 34 points, which tells you this landed as infrastructure rather than news. That is exactly how commoditization looks from the inside. If your product is the loop, the loop is now a free managed service from the company that also sells you the tokens. If your product is what the agent knows, what it is allowed to touch, and who is accountable when it goes wrong, today was a good day.
Comments