The 27-Minute AI Task: Custom Running Routes and the Transparency Problem

We have grown accustomed to artificial intelligence that operates in seconds—spitting out recipes, drafting emails, or summarizing articles almost before we finish typing our prompts. But what happens when you give an AI a task that takes nearly half an hour of autonomous work to complete?

A recent experiment with GPT-6 Astra running on ChatGPT Work highlights this fascinating shift from quick-reply chatbots to long-running autonomous agents. Developer Simon Willison gave the system a highly practical, real-world task: using OpenStreetMap data, design custom 5K and 10K running loops that start and end exactly at his front door.

Instead of generating a quick, generic list of nearby parks, the AI rolled up its digital sleeves and went to work for 27 minutes. It autonomously queried geographic databases to locate the address, downloaded local trail and road data, and calculated the optimal loops. The final output was remarkably polished: it provided downloadable GPX and GeoJSON files ready to be synced to a sports watch, alongside a custom-built, interactive map embedded directly into the chat interface.

It was a stunning display of what next-generation AI agents can do. However, the impressive result came with a frustrating downside: a complete lack of transparency.

The interface obscured the actual Python code the AI was writing and executing to solve the routing problem. By the time Willison realized he wanted to inspect the code to see exactly how the AI had pulled off the feat, it was too late. The system's memory management protocol—a process known as "compaction" that condenses chat history to save context space—had already wiped the step-by-step code from the record.

This scenario perfectly illustrates the double-edged sword of advanced AI systems. As they become highly capable digital assistants entrusted with intricate, multi-step workflows, they also risk becoming impenetrable black boxes. If an AI can work independently for 27 minutes to produce a result, users need the ability to audit its methods. Balancing this incredible new autonomy with user transparency will be one of the defining challenges as AI continues to integrate into our daily routines.

Key Points

• GPT-6 Astra demonstrated the ability to autonomously plan custom 5K and 10K running routes using OpenStreetMap data.
• The AI processed the complex geographic task for 27 minutes, generating both downloadable fitness files and interactive maps.
• The system's interface hid the underlying Python code from the user.
• Memory 'compaction' permanently erased the AI's working steps, highlighting a critical flaw in transparency for autonomous agents.

Why It Matters

As AI systems evolve to execute long-running, multi-step tasks autonomously, balancing user convenience with the ability to audit the AI's underlying process becomes a critical design challenge.

Sources:

• Generating running routes with GPT-6 Astra and ChatGPT Work — Simon Willison's Weblog