ChatGPT Work: 27-minute agent tasks reveal audit risks

In an evaluation published by technologist Simon Willison on his blog simonwillison.net, ChatGPT Work paired with the GPT-6 Astra model was assigned a complex geospatial task: identifying 5 km and 10 km running loops from a specific address using OpenStreetMap data. The system ran autonomously for 27 minutes before returning an interactive D3 map visualization along with downloadable GPX and GeoJSON files.

What does a 27-minute execution time mean for AI workflows?

A continuous processing time of 27 minutes marks a fundamental transition in how professionals interact with language models. Traditional AI assistance relies on immediate request-response loops that resolve in seconds. In contrast, long-horizon task execution allows an agent to query external APIs like Nominatim and Overpass, perform local geometric calculations, and construct web assets without step-by-step user intervention.

From a practical perspective, this transforms the model from a real-time conversational assistant into an asynchronous background worker. Instead of waiting at the screen, users trigger long jobs and collect finished multi-file deliverables later. However, extended runtimes require teams to re-evaluate how they monitor automated background execution.

Why does context compaction threaten code auditability?

The most critical finding from Willison’s experiment occurred after the task finished. When asked to inspect the exact Python code used to query the APIs and calculate the geographic loops, ChatGPT Work was unable to retrieve it. The platform had applied context compaction—a mechanism designed to trim conversation histories—which permanently removed the underlying code executed in the background sandbox.

This behavior requires separating what the source observed from the operational conclusions for technical workflows:

• Source observation: The AI successfully generated the final HTML and spatial data files, but the intermediate code executed during the 27-minute window vanished from the thread history.
• Practical consequence: Any enterprise workflow requiring compliance, security audits, or code inspection cannot rely on default conversation memory. If an agent executes scripts without explicitly writing them to persistent storage, the execution logic is lost forever.

How can professionals prevent data loss in long agent runs?

To safeguard technical tasks in environments using context compaction, operators must adapt their prompt strategies immediately rather than assuming history will be preserved.

Apply these operational guidelines for agentic automation:

1. Mandate persistent storage: Explicitly instruct the agent to save all generated scripts and code into named files within the persistent workspace (such as /workspace/script.py) before completing the task.
2. Extract structured logs: Request dedicated log files capturing API queries, schemas, and processing parameters alongside final visual outputs.
3. Separate code generation from execution: For mission-critical tasks, split the workflow into two distinct steps: generating and auditing the script first, then executing it in a second phase.