mckinsey/agents-at-scale-marketplace
Agents at Scale Marketplace for ARK (Agentic Runtime for Kubernetes)
What's novel
Agents at Scale Marketplace for ARK (Agentic Runtime for Kubernetes)
Code Analysis
12 files read · 3 roundsA Kubernetes-native marketplace of Helm charts, DevSpace configs, and Python services that deploy AI agent infrastructure — including a sandboxed code-execution operator with warm pools, MCP tool servers for domain-specific tasks (KYC, PDF extraction, web research), and pluggable LLM executors — all
Strengths
The ark-sandbox service is genuinely well-built: a kopf-based Kubernetes operator with a proper state machine (Pending→Running→Terminated), TTL-based expiry, warm pool management, and a concurrent MCP server — all backed by thorough, edge-case-rich tests. The MCP servers are practical and domain-specific (e.g., the Companies House resolver with name normalization and disambiguation, the PDF ownership extractor with two-pass relevance scoring), not generic boilerplate.
Weaknesses
The project is primarily a packaging/deployment layer (Helm charts, DevSpace configs, Dockerfiles) around services that are individually moderate in depth; the executors (claude-agent-sdk, langchain, openai-responses) could not be fully evaluated but appear to be thin adapters over existing SDKs. Some MCP servers (e.g., pdf-extraction) have hardcoded entity names in their relevance-scoring logic, and the pool controller's PVC-attachment path is explicitly a stub that only logs a warning.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
80
Contributors
10
Files
310
Active weeks
13
Repository
Language
Python
Stars
13
Forks
10
License
NOASSERTION