Building an Open Engine for AI Agent Handoffs

The Bottleneck is the Handoff

Most AI productivity issues stem from the 'hallway' problem, where a human acts as the manual bridge between isolated agent loops. Because agents live in separate chat windows, context is trapped, and the human must manually copy-paste information between tools like Claude, ChatGPT, and Codex. The breakthrough is shifting from 'prompt mode' (asking for an answer) to 'work mode' (assigning a task with clear state and definition of done) using a shared queue that acts as a system of record.

Implementing the Open Engine

The Open Engine framework uses a shared ticketing system (e.g., Linear, Jira) to coordinate work across different AI models. The system relies on five core components:

• Shared Queue: A central board where both humans and agents can read and write tasks.
• Standardized Task Structure: Each ticket must include the background, owner, agent instructions, and a clear 'definition of done'.
• Agent Skills: Specific protocols that teach the AI how to interact with the ticketing system, including setup, status updates, and task execution.
• Claim and Receipt Logic: Agents must 'claim' a task (moving it to an 'Agent Working' status) and leave a 'receipt' (the output and proof of completion) when finished.
• Smoke Test: A basic 'say hello' task that verifies the agent can successfully create, claim, and complete a ticket within the queue.

Moving from Chat to Work

To move from prompt-based interaction to a functional agent system, tasks must be self-contained. Instead of asking an agent to 'summarize a ticket' in a chat, a 'work mode' request involves classifying the ticket, attaching customer history, identifying known issues, and creating a follow-up task only if specific escalation rules are met. This allows agents to hand off work to other agents or humans without losing state, effectively removing the human from the middle of the loop.