Back to challenges
Challenge

KV-Cache Capacity Planner and Admission Controller with OpenAI Agents SDK and Claude 4 Sonnet

Prevent out-of-memory crashes on hardware nodes during long-context workloads. Build an admission control agent using OpenAI Agents SDK with Claude 4 Sonnet model capabilities to predict KV-cache allocation and reject or queue incoming requests exceeding physical VRAM safety thresholds.

EngineeringHosted by Vera
Sign in free to start
Challenge brief

What you are building

The core problem, expected build, and operating context for this challenge.

Develop an agentic admission controller using OpenAI Agents SDK and Claude 4 Sonnet specifications that dynamically measures dynamic KV-cache requirements and admits requests safely.

Delivery guide

How work is evaluated

Evaluation

Evaluates KV-cache footprint calculation accuracy and admission control shedding decisions.

Datasets

Shared data for this challenge

Review public datasets and any private uploads tied to your build.

Evaluation rubric

How submissions are scored

These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.

Dimensions
2 scoring checks
Binary
2 pass or fail dimensions
Ordinal
0 scaled dimensions
Dimension 1

OOM Safeguard Test

Ensure requests exceeding VRAM headroom are rejected or queued.

Binary check

This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.

Dimension 2

Admission Decision Precision

Accuracy of admitting maximal requests without OOM. • target: 0.98 • range: 0-1

Binary check

This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.

Learning goals

What you should walk away with

  • Implement a capacity prediction tool using the OpenAI Agents SDK.

  • Calculate dynamic KV-cache footprint based on sequence length and model layer depth.

  • Design safe request shedding and queuing algorithms for hardware clusters.

Resources and assets

Reference links and supporting material

Dataset notes

Synthetic batch of 50 long-context request configurations with varying layer dimensions and prompt lengths.

How this agent runs

Evaluates KV-cache footprint calculation accuracy and admission control shedding decisions.

Challenge input

JSON containing current_free_vram_gb, sequence_length, model_config

OpenAI Agents SDK

Official framework for building function-calling autonomous agents.

Claude 4 Sonnet

Frontier model context target for capacity planning.

Evaluated output

JSON decision (admit|queue|reject), calculated_kv_cache_mb

Checks for
  • Ensure requests exceeding VRAM headroom are rejected or queued.
  • Accuracy of admitting maximal requests without OOM. • target: 0.98 • range: 0-1
Proof of success
  • Benchmark: VRAM Management and Admission Control Standard
  • Admission Decision Precision target: 0.98
  • 1 public reference case
Runtime evidence
  • Python execution harness
  • Python sandbox (unavailable on Versalist)
View technical recipe

Configured tools

Action Space
  • OpenAI · Required
  • Claude Code · Optional
Observation
  • Claude Code · Optional
Policy Serving
  • Claude 4 Sonnet · Optional

Evaluation contract

  • OOM Safeguard Test · Weight 1
  • Admission Decision Precision · Weight 1

Recipe state

This is a preview. The configuration can change before the evaluation recipe is locked.

Frequently Asked Questions about KV-Cache Capacity Planner and Admission Controller with OpenAI Agents SDK and Claude 4 Sonnet