Latency

A coworker that takes too long to answer isn't used. Crewdesk measures where each turn's time goes and works to keep it down.

Measured per turn

Each turn records its phases: building the context, waiting for the brain, model calls, Hub calls, client calls, verification. The server's operator can see the time by phase and by route, with a target of 8 seconds at the median and 30 seconds at the 95th percentile, and an alert when it slips.

What keeps it fast

• Fewer model calls per turn, and a cheaper retry when the verifier sends a reply back.
• Tools with their own schema, sent to the model as they are, rather than in a strict mode that costs time.
• A kept connection to AMP. The Hub talks to the brain over one connection per chat, kept open for a few minutes between turns.
• A session ready ahead. There is always an empty brain session waiting for your next new chat.
• Warm helpers. The research, file analysis and CrewAI expert agents are kept warm.
• Small context. Project instructions are sent once and then only by their hash; earlier tool results are shortened as a long turn goes on.

What's left

Each round trip to AMP has a fixed cost that Crewdesk can't reduce with what AMP offers today, so the remaining work is on the number of round trips per turn.