Conversations¶
Most of the time, my agents work independently. You file a bead, I dispatch it, the agent handles it. But sometimes you want to think through a problem together with one of my agents. That's what pair mode is for.
Pair Programming¶
Pair-programming mode is a real-time conversation between you and an agent, scoped to a specific bead. The agent has full context of the bead and can execute actions (reading files, writing code, searching the codebase) while you talk.
To start:
- Open a bead in the Bead Viewer
- Click Pair
- Pick an agent from the dropdown
- Start talking
The agent responds via streaming -- you'll see the response build in real-time.
When to Use This¶
- Before work starts -- When requirements are fuzzy and you want to hash them out before the agent goes off on its own
- During implementation -- When you have opinions about the approach and want to collaborate rather than delegate
- During review -- When you're looking at what an agent did and want to discuss changes
- When debugging -- Two heads (even if one is artificial) are better than one
Persistence¶
I save the chat history for each bead. Come back later and the conversation is still there. You don't lose context between sessions.
API¶
# Start a pair session (returns a Server-Sent Events stream)
curl -N -X POST http://localhost:8080/api/v1/pair \
-H "Content-Type: application/json" \
-d '{
"bead_id": "bead-123",
"agent_id": "agent-456",
"message": "How should we approach the auth middleware?"
}'
# See all conversations for a project
curl http://localhost:8080/api/v1/conversations?project_id=my-project