agent-orchestrator

A daemon that lets me spawn purpose-built Claude agents from harness templates and watch them work in real time. The agents live in their own directories with their own CLAUDE.md, get evaluated and audited automatically, and can resume across multi-turn sessions.

What it does

• Defines agents as a CLAUDE.md template + tools + eval harness — a tight context bubble per task.
• Spawns and supervises them as long-running processes, not one-shot subagent calls.
• Streams live state to a TUI dashboard at ~500ms refresh so you can see what each agent is doing right now.
• Auto-evaluates outputs and audits behavior against the agent's spec.
• Supports session resume so an agent can pick back up across context boundaries.

Status

• Built and working as of 2026-03-20.
• Public mirror: LucianoLupo/agent-orchestrator-public (main repo private).
• Companion TUI dashboard at ~/projects/claude-monitor.

Agents I've shipped on it

• frontend-auditor — UI review agent, eval score 7.5
• cli-tui-generator — boilerplate generator for terminal UIs, eval score 8.5
• claude-monitor-research — research agent that produced design docs for the monitor itself

Why I built it

Claude Code's built-in subagent tool is great for one-shot delegation but doesn't fit "always-on worker" scenarios — long-running monitors, multi-day refactors, agents that need to keep state. Spawning agents as supervised processes with their own context windows is the missing primitive.

Connection points

• The harness pattern (CLAUDE.md per agent + tools + eval gate) is exactly what the harness-engineering research argues is the actual unit of leverage in agent work — this daemon is that synthesis turned into a runtime.
• claude-autoresearch runs as one of the agent types this orchestrator supervises.

Origin: Karpathy autoresearch + harness-engineering research from my Obsidian vault, distilled into a real always-alive daemon.