Agentic Workload Admission
Agentic Workload Admission
Agentic Workload Admission decides whether work may start. In this category, agentic names a workload shape: bursty arrivals, fan-out, and non-deterministic units. CI and agent sessions sit under one policy, including work that never calls a model.
Quelor admits work against machine resources: API budgets, executor slots, model credit windows, dollar budgets, dependency health. Resource classes are extensible by profile.
The lifecycle in one paragraph
AcquireLease runs before spend and returns ADMIT, DEFER(until, reason), SHED, or DENY. Unknown readings are never treated as available. RenewLease grants the next quantum for a long session; refusal means the holder checkpoints and hibernates. SettleLease is idempotent by lease id and returns declared cost against consumed cost. Classification is deterministic: resource-exhausted, dependency-unavailable, flaky, real-defect, or unknown. Exhaustion and outages exit degraded. Red is reserved for a real defect. Parked work waits on events — a window reset, health recovery, or lease expiry — and every verdict carries a decision_ref.
Adjacent categories
Each row is a system that already owns a real part of the path from intent to a finished run.
| Category | What they cover | What they miss |
|---|---|---|
| Durable execution (Temporal, Inngest, Hatchet, Trigger.dev, Restate) | Crash-safe workflows, retries, queues, and waits on events. | No first-class multi-resource admit-before-start across LLM budgets, executor slots, and dependency health. No CI and agent failure taxonomy, and no alert gating on that taxonomy. |
| CI platforms (Buildkite, Harness, CircleCI) | Runners, concurrency caps, and pipelines. | Capacity is a knob on the plan. It is not policy over API credits or model windows. |
| CI analytics and flaky tests | Flake detection, quarantine, and cost visibility. | The signal is post-hoc. These tools do not refuse work before it starts. |
| Merge queues | Serialized and batched merges, and speculative CI. | The scope is the forge. Agent sessions and LLM-budget admission are outside it. |
| LLM gateways and budgets (Portkey, LiteLLM, Helicone) | Per-request RPM and TPM, dollar budgets, and routing. | No deferred release on health events across CI and agents. |
| Agent orchestration | Task graphs, tracing, and hosted runs. | Multi-resource admission is limited. |
| Kubernetes admission and quota (Kueue) | Job-level admit-or-wait on CPU, GPU, and memory. | Cluster resources only. |
| AI FinOps | Showback, budgets, forecasting, and caps. | Reporting and spend limits. Not a control plane that defers work and classifies the outcome. |
Where Quelor sits
Quelor complements these tools. Durable engines still own crash-safe execution. CI platforms still own pipelines. Gateways still price and route a call. Queues, cluster quota, and FinOps views still do their jobs.
Quelor is the decision they call before spend starts. It admits, defers, sheds, or denies against the machine resources in the envelope, renews the lease for as long as the work is honest, and classifies the outcome when the lease settles.