Compression Levels

A compression level sets how much compression is applied to each request before it reaches the model. Higher levels cut cost more, but can affect accuracy in some cases.

There are four levels: High, Medium, Low, and None. Every agent selects one, and new agents start at High. The setting lives on the Agents page in app.induction.ai , under the Compression level column. You choose it when you add an agent, and you can change it at any time.

A change applies to new threads. A thread already running keeps the level it started with, so changing the setting never shifts an agent mid-task.

Choosing a level

Start at High and assess the results.

If accuracy holds, you are done. If it slips, step down to Medium, then Low. Each step down sends more of the original context to the model and reduces cost less. None turns compression off, which is also the cleanest way to measure what the other levels are doing for you.

How far down you land depends on the work. The more exact the model needs what you send, and the more exact you need what it returns, the less compression you should use.

The dashboard records the compression level each request ran under, next to that request’s cost and savings, so a change of level stays visible in the history.

Latency and accuracy

Compression is work, so it adds latency. We have seen it add about 10%, and the higher the level the more it adds.

Pass rates hold. On public benchmarks, High reaches the same pass rate as sending everything, and every level below it changes less still. The Toolathlon results show pass rates measured against the same agent running direct.

What to measure

Cost and cost savings are both on the dashboard. Success is yours to define, because only you know what a good outcome looks like for your agent, so use whatever check you already trust: a test suite, a grader, or a sample you read by hand.

Watch the whole thread, not the single request. Induction works best over long threads. The first few turns may not save anything, but a smaller context carries over into every turn that follows, so the savings build as the thread grows.