AgentManager

AgentManager

Conduct autonomous agents at scale safely. You lead. Agents execute. Human-on-the-loop, NOT human-in-the-loop. Orchestrate AI work like a manager, not a prompt juggler.

Watch the demo on YouTube

What this actually is

AgentManager runs Claude Code CLI processes. Each agent is a real Claude Code session with full tool access (file editing, bash, git, MCP integrations) running in an isolated workspace. The platform handles the operational side:

- Agents persist across restarts. State syncs to GCS. Cloud Run scales to zero, agents resume on wake.
- 6-layer emergency kill switch. Process kill -> token rotation -> GCS remote kill. Built because I needed it.
- cgroup v2 memory monitoring. Tracks container memory via cgroup, rejects new agents at 85%. No OOM surprises.
- Process group lifecycle management. Negative PID signals kill entire process trees. No orphaned processes.
- Git worktree GC. Persistent bare repos with per-agent worktrees. Automatic cleanup on destroy, startup and every 10 minutes.
- Inter-agent message bus. Task delegation, results, questions, interrupts. Auto-delivery to idle agents with delivery locks to prevent races.

Not a wrapper around the Anthropic SDK. Not a chat UI. It runs actual Claude Code processes with the guardrails you need when agents start doing things you didn't ask for.

Why Claude Code CLI, not the Anthropic SDK?

The SDK gives you chat completions with tool calling. Claude Code gives you a complete coding agent with file editing, bash execution, git, MCP, session resumption and sub-agent delegation, all maintained by Anthropic. AgentManager runs these agents rather than trying to rebuild them from scratch.

Claude Code's --output-format stream-json gives typed JSON events (not terminal scraping) that the platform…