Abstract
A fleet of autonomous LLM coding/review agents, each holding its own per-seat license, draws from a small set of shared provider quotas (an organization spend ceiling and a per-minute API rate limit). Treating every provider error as a uniform retry-with-backoff is wrong twice: a transient rate limit and a durable capacity exhaustion need opposite recovery horizons (minutes vs. hours), and a central platform key cannot test whether an individual seat's license has recovered because the quota is bound to the seat's own credential. This publication discloses a method that (1) recovers a limit signal from the agent's free-text output — the agent is a CLI subprocess whose only return surface is text — and classifies it into a durable-capacity tier and a transient-rate tier with the rule transient = rateLimited AND NOT hardCapacity, escalating to durable when both classes co-occur; (2) applies a tier-specific cooldown horizon; (3) reroutes the in-flight task to the first non-cooled healthy peer in the role-appropriate pool via an ordinality-preserving pick, leaving the task open (never lost) if none exists; and (4) recovers a cooled seat only when an out-of-band active probe that exercises that seat's own per-seat credential under a hard abort timeout returns a non-limit response. We present architecture, a classifier state machine, reroute and probe sequence diagrams, a relational data model, a clean-room Node.js reference implementation, a prior-art delta against the 2026 industry literature, and one independent plus fifteen dependent claims.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Assuncao, gustavo matthew, "Two-Tier Seat-Limit Classifier with Active Per-Seat-Credential Self-Probe", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10858