Cursor-like personal AGI

Classify chat vs agent, route the best model per turn, and only run plan-then-act for complex work.

Cursor-like personal AGI

We’ve overhauled how the personal agent loop works so it behaves less like a rigid tool-hammer and more like the conversational, context-aware assistants you’re used to (like Cursor).

When you call /step:

1. Conversation: Normal talk, Q&A, and clarifications are now answered with a conversational reply. The agent no longer forces a plan/tool loop for simple chat.
2. Model Router: The system classifies each turn (chat, ask, or agent) and picks the best model from your preferences. Fast/cheap models handle chat, while stronger models tackle complex agent work. You can steer this with optimizeFor (cost, balance, or intelligence).
3. Agent Loop Only When Needed: Multi-step goals and complex side-effects still get the full plan → think → act → observe treatment.

Responses from /step now include a phase (chat, ask, planned, acting, done) and a routing block (intent, complexity, model, optimizeFor), so clients can show “used Model X” dynamically.