Context Management

On every request, Induction decides what the model needs to see on this step and reshapes the conversation before it reaches the provider. Four techniques do the work.

Semantic compression

Tool results are the bulk of an agent’s context, and most of a large tool result is padding around the part the agent uses. Induction extracts the relevant information from large structured and unstructured tool output, so the signal stays and the padding stops being re-sent on every subsequent call.

Context lifecycle

A tool result matters for a few turns, then usually never again, but it keeps getting re-sent anyway. Induction classifies tool context as relevant or stale, archives what has gone stale, and restores it if the agent turns out to need it. A long-running agent stops dragging its whole history through every request.

Duplicate elimination

Agents re-read files, re-run searches, and re-fetch the same records. When a tool result repeats one already in the conversation, Induction replaces the repeat with a lightweight reference, but only when doing so saves tokens.

Economic optimization

Every technique above has a cost: some spend model calls to decide what to keep, and rewriting a request can invalidate the provider’s prompt cache, which has a price of its own. Induction weighs model pricing, cache behavior, likely future turns, and its own overhead before changing a request. When the math does not favor touching it, the request goes through unchanged. That discipline is why the benchmark costs, which include all of Induction’s overhead, still come out 37–67% lower.