Breaking the Chain: Agent Continuations for Resumable AI Workflows
summary
Claim: an agent can capture its full call stack—tools, goals, partial responses—into a compact JSON blob combined with the messages array to form a protocol‑level Agent State that pauses, migrates, checkpoints, and restarts. Prototype is Python on the OpenAI API, with reference JSON and a live demo that suspends a three‑layer agent for human approval and successfully resumes with normal messages output. Works with nested subagents and aims to resume through standard LLM function‑calling APIs. Links on the page point to the repo.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:14General Challenges
- 3:01Basic Agent Execution
- 4:07Agent Execution Scenarios
- 5:56Agent Loop Persistence
- 8:12Agent Continuations
- 10:00Messages Array
- 11:33Continuation Extension
- 14:42Agent Continuation
- 17:22Continuation Object
- 18:47Example
- 21:45Demo
- 24:08Summary
related talks
26mBeyond Sandboxes: Architecting Durable Runtimes for AI Agents
27mSecure "Human in the Loop" Interactions for AI Agents
26mAgent Omelets & Other Forbidden Techniques
11mCreate AI Agents That Work Independently
34mAI Agents: Transforming Autonomous Decision-Making
17mYour Agents Need a Save Button
40mBuilding Effective AI Agents Simple, Composable Patterns
48mAI Coding Agents and how to code them