Three production agents serve live requests. The system observes every interaction, scores it across four dimensions, runs sandbox experiments in shadow mode, and proposes improvements — all without ever putting itself in the limelight, unless of course you want it to.
The collection
For blending in — or standing out
The intake agent takes its cue from the first contact: clear, conversational, unassuming. Refining and updating this approach with calibrated brevity, intake handles general questions and routes complex requests without ever putting itself in the limelight.
Anywhere, any depth
Performing equally well at the surface and in the deep, the research agent is a pleasure to work with. Use it as a solitary piece to light up a single question, or in clusters above larger topics to add a jolt of insight, colour, and accent.
Get to the point
The action agent executes specific tasks and returns structured, actionable output. Precise. If a task is ambiguous, it asks one clarifying question before proceeding — no posturing, no filler, just the work.
Method
Every chat is scored automatically across latency, output length, user feedback, and error rate.
Experimental sandbox agents receive the same live traffic but never respond. Pure observation.
Every night the autoresearch loop reads 7 days of data and proposes improved prompts via LLM.
A four-check gate (traces, delta, error floor, latency) advises. Only a human pressing approve makes it live.