Problems / AI systems, hard

MCP agent platform

4 tasks/s. The agent thinks (model) and acts (tools through an MCP gateway) 6 times per task, with vector memory and results in S3.

Free to open. Runs in your browser, no sign-up.

Your goals

Keep success ≥ 95% and p99 ≤ 60 s. A 1% tool error compounds over the steps: find it and cut it, and watch the cost.

The traffic

4 requests a second for 60 seconds: 100% task (writes).

The system

What each component does in this design, and what calls it. The simulator mimics each one from these settings.

ComponentIts job hereSettings
Load balancerLoad balancerTraffic enters here.Calls Agent API.Layer 4 balancers route by IP and port; layer 7 balancers read HTTP and can route by path or header.–
Agent APIApp serviceLoad balancer forwards to it.Calls Agent orchestrator, S3 (task results).Little's law: throughput ≈ concurrency ÷ latency. 200 slots at 50 ms serve about 4,000 req/s.2 replicas × 400 slots · timeout 120 s
Agent orchestratorAgent orchestratorAgent API calls it.Calls Agent memory (vector), Model, MCP gateway (auth, tool registry).An agent loops: think, call a tool, read the result, repeat until done.300 slots · 6 steps
ModelLLM endpointAgent orchestrator calls it.Latency has two parts: time to the first token, then a steady stream of tokens per second.200 streams · 1,200 calls/min
MCP gateway (auth, tool registry)MCP gatewayAgent orchestrator calls it.Calls MCP tool servers.MCP (Model Context Protocol) is an open standard for connecting models to tools and data.500 slots · 30,000 calls/min
MCP tool serversTool / MCP serverMCP gateway (auth, tool registry) calls it.Tools give the model actions such as search or database queries; each call is another round trip.40 slots
Agent memory (vector)Agent memory storeAgent orchestrator calls it.Agent memory keeps facts across sessions so the agent does not start from scratch.500 slots
S3 (task results)Object storageAgent API calls it.Object stores like S3 keep copies across devices and facilities for very high durability.5000 slots · 300,000 calls/min

More ai systems problems