Separate the loop from its state and tools
The model proposes actions; the surrounding software supplies context, routes calls and feeds results into the next iteration. Anthropic’s April 2026 account separates that harness loop from a durable session log and the sandbox that performs work. Other implementations draw these boundaries differently. A saved session is not the same object as the model’s current context window or the external service’s records.
In episode 69 at 15:29, Joel Hron describes a simple harness exposing Thomson Reuters applications during training. The harness supplies access; training changes how the model uses it.
Diagnose a restart before repeating an action
Fictional example: an AI agent requests a document-conversion job using an application operation ID, J17. The worker stops before it saves the tool response. The request might already have reached the conversion service. A missing local response does not prove the external operation failed.
| Record | Status | Next step |
|---|---|---|
| Local request only | Unknown | Ask service for J17 |
| J17 accepted | Not confirmed complete | Check job status |
| J17 completed | Recorded complete | Check output artifact |
| No service status | Unknown | Flag for review |
An operation ID helps correlate records; it prevents duplicate work only if the service actually supports and enforces that behavior. Inspect the service contract before retrying a write.
LangGraph’s checkpointer documentation explicitly says replay runs nodes after a selected checkpoint again, including API calls. Its durability modes also differ in when state is written. A generic claim that an agent “has checkpoints” does not tell you which external effects can repeat.
Test what survives an interruption
In a controlled environment, interrupt before a tool call, after the service accepts it and after the result is durably recorded. Inspect the session, execution state and external result separately. Test an unavailable status service too; an unresolved outcome should remain visible rather than becoming a confident completion message.
AI observability connects recorded calls and results for diagnosis. If the tool returns an authorization error, inspect access and execution controls; repeated wrong choices with working access point to different questions about instructions, context and the model. Use the framework decision guide to compare implementations, and LLMOps to track the application versions involved in the failure.
Hear it from the guest
“We will put the model into a harness, like a very simple agentic harness, with access to our TR applications and tools.”
Quotes lightly edited to remove filler words.
Sources
- Anthropic: Scaling Managed Agents — April 2026 account that separates the harness (the loop that calls the model and routes tool calls), the append-only session log and the sandbox; checked November 2, 2026.
- LangChain: LangGraph checkpointers — Replay re-executes nodes after the selected checkpoint, including LLM calls and API requests; the exit, async and sync durability modes set when state is written; checked November 2, 2026.
Go deeper
- Do you still need an AI agent framework? AI, decoded · Do You Still Need an AI Agent Framework?
- How do AI coding agents like Claude Code work? AI, decoded · How AI Coding Agents Work
- Anthropic: Effective harnesses for long-running agents article
Study progress records, environment setup and checks that help an agent resume work across sessions.
From the conversation
-
Thomson 1: The New $40M Legal AI Model | Thomson Reuters Joel Hron -
Agent Memory: The Last Battleground in the AI Stack | Richmond Alake, Oracle -
The AI Framework Era Is Over: Why Context Is the Moat | Jerry Liu of LlamaIndex -
For AI Agent Fleets, the Model Is Rarely the Bottleneck | Theory Ventures' Andy Triedman -
Why Context Alone Isn't Enough for Enterprise AI Agents | WisdomAI CPO Kapil Chhabra -
AI Codes: Product Engineers Decide What to Build | Laurie Voss, Arize -
Genspark's Bet: AI Agents Become the Users of Software | Wen Sang