example: Deploy Claude Code as a Kagenti-managed agent

[jenna-winkler]

Description

Deploy a Claude Code container on Kagenti and validate that the platform's core capabilities (identity and governed tool access) work with a harness-style agent.

Today all Kagenti examples are framework-based agents (e.g., LangGraph) that follow a build-deploy-discover lifecycle. Claude Code is different - there's no build step, no A2A agent card, no framework code. You configure it with a system prompt, skills, and tools. This example validates whether Kagenti's platform value applies to that model.

What the example demonstrates

• Credentials - Anthropic API key stored as a K8s Secret, injected via env var.
• Identity (SPIFFE/SPIRE) - Claude Code pod gets a SPIFFE workload identity like any other Kagenti agent.
• Tool access (MCP Gateway) - Claude Code connects to platform-managed MCP tools through the MCP Gateway.
• User interaction - Users talk to the agent via Slack/Discord bot.
• Session persistence - Agent home directory backed by a PVC so Claude Code's native session storage survives pod restarts.

Out of scope (follow-up)

• Observability (Phoenix tracing for harness sessions)
• UI integration (agent visible and testable from Kagenti dashboard)
• Scheduling and triggers (cron, webhook, GitHub Actions)

[jimcadden]

@esnible this seems aligned with what we spoke about today. Can you pull in Rong and Jarek too?