We thought people would build their agents on our platform. It turned out they wanted to build them their own way and just hand us the running.

The first version of our platform asked people to move in. It had pods for teams of agents, workflows to chain them, a memory bank, and our own SDKs in three languages. The idea was that agents would be built inside our world. Then we watched what people actually ran. Most had a framework they already liked, or a loop they'd written themselves. Some were running agents as a pile of scripts and CLI processes, and hated it. Nobody wanted a nicer way to design an agent. They wanted someone else to keep it running. We'd built a whole kitchen, and people mostly wanted the oven. So 2.0 dropped pods, workflows, the memory bank and the SDKs. The platform now speaks the same API as OpenAI's Agents API. That's only the shape of the calls. The model behind them can be Claude, Gemini, an open-weight model, or anything on OpenRouter or on the user's own key. Anyone already on the openai package changes the base URL and nothing else (the line is in the comments, since posts with links get removed here). Everyone else has an OpenAI client in almost every language, far more than we could ever have maintained. Leaving means changing the URL back. What's left is what people kept asking about: what happens when an agent breaks, what it cost, and whether it ran the model they asked for. When it breaks. Turns fail. A provider has a bad minute, we restart a machine, the credit runs out. We made it so the turn fails and the session survives. Saying "continue" picks it up with its conversation and files intact, without rebuilding context or rerunning the last hour. What it cost. Every model call is logged with its cost, down to a millionth of a dollar, so "what did that run cost?" has an answer without an export. Which model ran. The model someone picks is the model that runs. If it's unavailable, the call fails with an error. Some gateways quietly fall back to a different model, and people find out when the outputs start looking off. We'd rather fail. Under every session is the same open-source harness from our post here a few weeks back, the one where 100 agents updated our docs (link in the comments). It was built to use as few tokens as possible, and that's still its main job. And since someone always asks: we make nothing on tokens. On our credit, people pay OpenRouter's price plus the fee OpenRouter charges for buying credit. On their own key, they pay their provider directly. We charge for plans and machine time. For anyone running agents in production: what did you end up building yourself that you wish you hadn't?