Fewer wasted tool calls
The agent's small mistakes with its tools now get fixed or explained in one step, based on what thousands of real calls showed.
We looked through more than three thousand of the agent's recent tool calls to find where time and space were lost. When the agent writes a request with a small slip, such as a line break or a stray backslash inside some text, the request is now tidied and carried out instead of refused. When it asks for a file that is not there, it is told the nearest names in that folder, so it can fix the path on the next try. Opening a file that is not text, such as an image or an archive, now gives its size instead of filling the conversation with unreadable bytes. When several changes to one file are sent together and some do not fit, the agent hears about all of them at once.
Scripts that run many steps
When the agent runs a short script that does many steps at once, the script has two minutes in total. Builds and test runs belong outside it, and the agent is now told so. A script that runs out of time is no longer simply retried the same way. And when one step in a script fails, the agent hears which step it was and how to keep the other steps going.