Month-end readiness for an agent fleet

A complete build of the capstone in Python. Two mock Claude platforms and a mock Batches API, a Redis Lua limiter shared by four worker processes, per-firm deficit round robin, one retry layer with budgets and breakers, priority admission with aging and a batch path, an exact response cache, a sticky router with a token-metered move budget, virtual-clock checks for every milestone, and a Locust month-end replay ending in a capacity plan.

Back to the project

The dispatch layer is a pipeline every model step passes through in the same order: admission (class, aging, intake bound, latest start), fair queue (per-firm deficit round robin inside each class), cache (only for document-keyed sub-tasks), router (sticky platform per matter), limiter (shared Redis buckets) and concurrency cap (AIMD), then the call with one retry layer and a breaker per platform. The decision that matters most is that the pure parts (limiter algorithm, fair queue, admission, retry decision, breaker, router) take an injectable clock, so the same classes run in the live fleet and in a virtual-clock simulator: hours of month-end are checked in seconds, and the live load test only has to confirm what a simulation can't see.

Back to the project

Month-end readiness for an agent fleet

Month-end readiness for an agent fleet

Sign in to keep reading the answer.

Sign in to keep reading the answer.