Induction

Induction is an agent optimization layer that sits between your agents and the models they call. One base-URL change points your existing OpenAI SDK at it, and from then on it manages the context each request carries. Same models, same success rates, zero code changes beyond that line. It serves the Responses API (Responses only, not Chat Completions), for OpenAI models and for open models hosted on Fireworks.

Induction is built for long-running agent workloads: agents that work through dozens or hundreds of tool-calling turns per task. It is not targeted at coding agents.

The challenge of context windows

As an agent works, its context fills with information that is no longer relevant but is still sent to the model on every step: instructions, tool schemas, the full history, every tool result so far. If each call adds ~20K tokens of new context, cost grows with the square of task length:

That is one task. Most of those tokens are the same context sent again and again, and most of it no longer matters to the next step.

Caching and right-sized models reduce the cost, but they do not eliminate the underlying waste. Summarizing history loses details the agent turns out to need, and capping turns or tools limits what your agent can do. As your agents take on longer workflows, the waste compounds.

Make every token count

Induction sits on the request path, where the whole conversation is visible on every call. It learns your workloads, classifies relevant versus stale context, and removes redundancies in real time:

• Compress tool results: extract the relevant part of large structured and unstructured tool output.
• Archive stale context: set aside tool context the agent is done with, and restore it if it is needed again.
• Drop duplicates: replace repeated tool results with lightweight references when that saves tokens.
• Only when it pays off: every decision accounts for model pricing, cache behavior, likely future turns, and the cost of the optimization itself. If touching a request would not save money, it goes through unchanged.

The result is focused context that keeps wasted tokens low and success rates high. Context Management covers each technique in more depth.

What stays the same

You keep your SDK, your models, your prompts, your tools, and your provider API keys. Keys travel with each request, and Induction does not hold them. Responses come back in the provider’s native format. If a request cannot be optimized safely, it falls through to the provider unchanged.

Proven on agent benchmarks

On public agent benchmarks, the same agent with the same model completes the same tasks for 37–67% less. Pass rates hold at baseline in every configuration tested. At best, the same budget runs 3x more tasks. Numbers and setup are in Benchmarks.

Where to go next

• Quick Start: point an agent at Induction in one line.