KV-Cache Capacity Planner Agent Mesh with CrewAI and Cursor
Develop a KV-cache memory capacity planner using CrewAI and Cursor to optimize page allocations for concurrent 128k context requests. Achieve zero out-of-memory preemption events across a 1,000-request benchmark while maintaining p95 prompt processing latency below 250ms.
What you are building
Orchestrate role-playing CrewAI agents developed in Cursor to manage vLLM-style paged KV-cache allocation rules.
How work is evaluated
Evaluates OOM preemption rates and latency profiles of the KV-cache planner under continuous load.
Datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
Zero Preemptions
Ensure zero requests are preempted due to out-of-memory conditions.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
P95 TTFT Latency
95th percentile latency for time-to-first-token in milliseconds. • target: 200 • range: 0-250
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Learning goals
Implement multi-agent role delegation using CrewAI for capacity planning.
Design dynamic block-allocation strategies for continuous long-context KV caches.
Prevent GPU memory fragmentation and request preemption under burst loads.
Measure impact on context swapping latency and time-to-first-token (TTFT).
Resources
Trace dataset containing context lengths and arrival times for 1,000 synthetic agentic interactions.
Challenge setup
Evaluates OOM preemption rates and latency profiles of the KV-cache planner under continuous load.
Challenge input
JSON payload containing request arrival rates and context lengths
CrewAI
Orchestrates dedicated role-playing agents for capacity management
Cursor
Accelerates development and refactoring of complex system tools
Evaluated output
JSON containing preemption_count and p95_ttft_ms
- Ensure zero requests are preempted due to out-of-memory conditions.
- 95th percentile latency for time-to-first-token in milliseconds. • target: 200 • range: 0-250
- Benchmark: vLLM Memory Management Benchmark
- P95 TTFT Latency target: 200
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Shows the first task, up to three tools, and up to three evidence items.
View technical recipe
Configured tools
- CrewAI · Required
- Cursor · Optional
- crewAI · Optional
- Cursor · Optional
Evaluation contract
- Zero Preemptions · Weight 1
- P95 TTFT Latency · Weight 1
Metrics
- P95 TTFT Latency · Target 200
Benchmarks
- vLLM Memory Management Benchmark
- general
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON payload containing request arrival rates and context lengths
Output: JSON containing preemption_count and p95_ttft_ms
Execution harness: Python
Evaluation tools
- CrewAI · Agent Framework
- Cursor · AI Engineering Tooling
Recipe freeze marker
No recipe freeze time is recorded.