Who owns the job after the agent disconnects? A Kubernetes MCP architecture
stevenacreman · reddit · 2026-10-09
The author of KubeDEX published an article on using Kubernetes as infrastructure behind coding agents. The core MCP question: when an agent kicks off an expensive tool call (like requesting a test environment) and then disappears, who owns provisioning, test execution, and cleanup? This is a proposed architecture, not a production report.
Key recommendations:
- MCP should expose bounded platform operations with durable job IDs and status stored outside the conversation
- Cluster API manages cluster lifecycles, Argo CD reconciles applications, CI/workflow jobs run verification
- A separate controller should enforce environment leases and cleanup — Kubernetes Job TTL alone won't expire a namespace, cluster, or leftover cloud resources
- The result should be evidence tied to the commit, image, fixtures, and test run, retained after the environment is removed — the agent can request experiments but can't rewrite what counts as passing
The article also covers MCP implementations and isolation tradeoffs between namespaces, virtual clusters, and separate clusters.
More from coding & agent
- PartyKit shuts down free hosted platform 2.5 years after Cloudflare acquisition — threepointone · 2026-10-09
- Run a Brand X Account With 3 AI Bots: Community Manager, Content Strategist, Analyst — FinanceYF5 · 2026-10-09
- Google's A2A protocol moves to neutral foundation alongside MCP under AAIF — SnooDingos9560 · 2026-10-09
- LangChain Founder: Evals Work for Narrow Tasks but Break Down for Autonomous Agents — hwchase17 · 2026-10-09
- Deepkit author: Bun is rediscovering our decade-old ideas, agents could revive it — MarcJSchmidt · 2026-10-09
- Qdrant ships LangGraph memory store integration for vector-native agent memory — qdrant_engine · 2026-10-09