Cortex
The learning layer. Cortex watches what your agents actually do, finds the work that repeats, and learns a cheaper way to serve it.
Most of what an agent does, it has done before. Roughly three quarters of a typical enterprise workload is repetition, and that repetition carries most of the bill. Cortex is the part of Operant that notices.
What it watches
Every call through the Gateway is fingerprinted — a one-way hash of the request's shape, not its contents. Identical shapes cluster. When a cluster crosses a threshold, Cortex has found a workload worth learning.
What it learns
A recurring workload usually carries far more context than the task needs: a system prompt written for the general case, tool definitions that never fire, and the residue of turns that no longer matter. Cortex derives a skill — a tighter prompt, a smaller tool surface, and a model tier that is adequate rather than maximal.
What it costs you to let it learn
Learning needs the text. Once a workload recurs, Operant begins storing the content of that workload's messages, tool inputs and tool outputs, and backfills the earlier turns of the conversation. This is on by default, and it is the trade Cortex makes: capability for privacy. It is described precisely in section 5.3 of the Privacy Policy, and you can ask us to turn it off.
If that trade is not one you want to make on a given workload, do not route that workload through Operant, or ask us to disable escalated capture for your account.