Breaking the Chain: Agent Continuations for Resumable AI Workflows - Greg Benson

2.4K views · Jun 03, 2025 · 27:01 min · Watch on YouTube ↗

Takeaway

Agent continuations let you snapshot, persist and resume in-flight agent state across hosts so HITL approval and long-running workflows survive process restarts.

Summary

Benson (SnapLogic chief scientist, USF prof) introduces 'agent continuations' — borrow programming-language continuations to snapshot the entire state of a running agent (including sub-agents and LLM/tool call history) so it can pause and resume.
Solves two production pains: human-in-the-loop approval for high-stakes tool calls without keeping the agent loop running, and resilience for long-running agents to survive failures and rate-limit errors.
Key insight: the LLM messages array already logs interaction history; agent continuations build on top of it with the extra runtime state needed for full resume, so the loop process can fully shut down between resumes.
Multi-level agent configurations (orchestrator + sub-agents) are supported, and the framework uses tool decorators similar to other agent SDKs.

agentshuman-in-the-loopdurable-execution

Original description

AI agents are powerful—but brittle. Once an agent chain starts, you either let it run or you tear it down and lose state. Agent Continuations change that contract. Borrowing from programming‑language continuations, we capture an agent’s entire call stack—tools, goals, partial responses—in a compact JSON blob combined with the familiar messages array. The result is a protocol‑level "Agent State" that lets you:

- Pause anytime for human-in-the-loop approval gates, rate‑limit resets, or progressive UI updates.

- Migrate agents across nodes, clouds, even different agent execution platforms

- Checkpoint long‑running multi‑agent plans using off‑the‑shelf storage and enable restarting in the presence of agent failure

- Resume seamlessly through standard LLM function‑calling APIs, so every framework that speaks OpenAI JSON can speak continuations.

Our approach works with single-level agent loops and multi-level agents in which agents can call subagents.

Attendees will leave with open‑source Python snippets and a mental model that turns “monolithic” agents into restart‑able, human‑aware services—shrinking failure windows and unlocking new UX patterns for AI products.

**Key Takeaways**

- Why Continuations are a good construct for Agent State
- Protocol spec and reference JSON examples and a - Python implementation
Live demo: suspend a three‑layer agent with suspending for human approval

** Links **

https://github.com/SnapLogic/agent-continuations
https://agentcreator.com