Krate Agent Dispatch overview

End-to-end workflow for dispatching an AI coding agent within Krate. Triggered by a user command (UI button, kubectl create, PR comment /agent), a schedule (AgentSchedule), or a pipeline step. Steps: 1. Dispatch request: An AgentDispatchRun resource is created with the target repository, branch, prompt, and AgentStack reference. 2. Permission review: The platform evaluates the dispatch against org-level AgentPermissionSets and PolicyProfiles. If the requested tools or resource limits exceed policy, the run is rejected or escalated for org-admin approval. 3. Context bundle assembly: Krate checks out the repository at the specified branch, injects CLAUDE.md and relevant memory snapshots from the company brain, and prepares the agent workspace container. 4. Agent-mux launch: The dispatch controller creates a Kubernetes Job running the agent-mux container with the selected coding agent (Claude Code, Codex, etc.). The session streams output back via SSE. 5. Completion: On agent completion, results are collected — files modified, PR created/updated, tests run. A memory snapshot is generated and stored. The AgentDispatchRun status is updated to Completed or Failed.