An AI-driven agent framework where conversations become workflows.
This project is an experiment in building a persistent, multi-agent system that acts as a true operating system for thought. Unlike traditional chatbots that reset every time, this framework allows you to interact with a single Orchestrator agent that can spin up, manage, and recall specialized Execution agents—each with its own memory, tools, and responsibilities.
- You always talk to the Orchestrator.
- The Orchestrator decides whether to reuse an existing agent or spawn a new one.
- Execution agents persist, remember, and can be recalled at any time.
- Tools and tasks give agents real capabilities (drafting emails, scheduling reminders, searching knowledge).
- Memory layers (short-term, episodic, semantic) ensure the system scales without forgetting critical context.
Example:
User: Email Alice and Bob about lunch tomorrow
Orchestrator: Spawning two email agents...
Agent(Alice): Draft ready.
Agent(Bob): Draft ready.
Orchestrator: Both drafts are ready. Send?
Later…
User: Did they respond?
Orchestrator: Recalling the existing Alice and Bob agents…
Agent(Alice): No reply yet.
Agent(Bob): Confirmed availability at 12:30.
-
Persistent Agents: Execution agents with their own logs, memory, and context.
-
Orchestration Layer: A router that maintains UX consistency while delegating real work.
-
Memory Model:
- Short-term: recent messages.
- Episodic: rolling summaries with temporal anchors.
- Semantic: embeddings for durable facts and artifacts.
-
Tooling Framework: Agents can call tools (atomic ops) or tasks (multi-step routines).
-
Interfaces: Start with a Next.js chat UI; later extend to RCS/iMessage for natural messaging.
- Frontend: Next.js (React, streaming chat UI, Agent Dock, Memory Viewer).
- Backend: Convex (serverless DB + scheduler + actions).
- Auth: Clerk.
- LLM Orchestration: OpenAI / Anthropic with function-calling.
- Memory Store: Convex tables for short-term, episodic summaries, and semantic embeddings.
- Tools: Email drafting/sending, reminders, search.
- MVP: Chat UI + Orchestrator + Email draft tool + Reminder scheduling.
- Phase 2: Episodic & semantic memory layers, richer tools (notes, search).
- Phase 3: Multi-interface (RCS, iMessage), project workspaces, self-healing agents.
Most AI “assistants” today are stateless and forgetful. This project shows what’s possible when you treat AI as a system of agents with memory and roles, not just a single model call. The goal isn’t just to demo LLMs, but to build the foundation of an AI-native operating system for human thought and action.