Someone finally built the thing the AI lab demos promised but never shipped: agents that wait for permission.
The Summary
- Open Duet is a self-hosted framework for creating persistent AI agents with memory, workspace isolation, and approval gates before they act
- Alternative to Claude Duet, OpenAI Dots, and ChatGPT agents, but you control where it runs and which models power it
- Agents can delegate work to specialists, maintain context across sessions, run on schedules, and operate with three permission levels: Ask first, Accept edits, or Plan only
The Signal
The architecture matters more than the features. Open Duet separates the orchestration layer from the model layer. You bring your own API keys. The framework handles memory persistence, tool approval, workspace isolation, and agent coordination. This is the opposite of how Anthropic and OpenAI ship agent products: locked to their models, their infrastructure, their permission systems.
The permission model is the interesting part. Three modes: Ask first (agent proposes, you approve), Accept edits (agent acts but you review), Plan only (agent thinks but never executes). Most agent frameworks assume autonomy is the goal. Open Duet assumes humans want granular control over what runs when.
"You choose where the app runs and which engine gets your prompts."
The team delegation feature requires Claude right now, which undercuts the model-agnostic pitch. But the structure is sound: a lead agent parcels out work to specialist agents, each with their own instructions and workspace. The lead tracks progress, the specialists report back, you follow the whole thing in a chat interface. It's closer to how actual teams work than the solo-agent-does-everything pattern most frameworks default to.
Memory and routines push this past chatbot territory:
- Agents maintain notes across sessions and surface relevant context automatically
- Routines run on schedules or webhooks without human initiation
- Each agent gets workspace isolation so file operations don't collide
- Conversation history search lets agents reference earlier exchanges
The project is in active development. The GitHub readme flags known limitations and warns against using it for critical work. That's honest positioning. Most agent frameworks oversell readiness.
The Implication
If you're building internal tools and don't want your data flowing through Anthropic or OpenAI's agent APIs, this is worth testing. The approval gates make it viable for environments where "let the agent run loose" isn't an option. Watch how the delegation feature evolves. If they decouple it from Claude, this becomes a real alternative to the closed agent platforms. The self-hosted angle matters more as companies realize agent activity logs are just another dataset they'd rather not hand over.