constelio

Evidence from real runs

Agents that learn to need less AI

A Constelio fleet starts by reasoning with a language model. When a task keeps coming back, the meta-agent turns what works into plain code. Below are measurements from our own fleets, as they were recorded.

Case 01

Collecting server status: from 6 minutes to under a second

A collector agent gathers the status of five machines by asking the agent on each one. Same task, same fleet, a few hours apart.

MetricBeforeAfter
Duration6 min 42 s909 ms
Tokens in100,4600
Tokens out2,4080

≈ 440× faster

and no language-model cost at all

BeforeFirst run · reasoning

The collector uses its language model to plan the task, ask each agent in turn and assemble the answer. It works, but it is slow and spends tokens every time.

Timeline of the first run: a 6 minute 42 second session in which the collector makes a series of language-model calls and asks the observer, test02, test03, test04 and whisper agents one after another
AfterA few hours later · learned

The recurring pattern has become a deterministic run, collect-host-status. The same five agents are asked, by plain code, with no language model involved.

Timeline after learning: a single 909 millisecond run, collect-host-status, that asks the same five agents in sequence without any language-model call

From the call trace

Faster

From minutes to milliseconds for a task the fleet performs again and again.

Cheaper

Zero tokens means no inference cost and no GPU time for routine work.

Predictable and traced

The same steps every time, and every call still recorded: both timelines come from the fleet's own call tracing.

More cases will be added here as our fleets keep learning.