Two months ago, GPT-5.6 Sol caused a major incident.
It deleted nearly all of Matt Shumer’s files on his Mac, including the company’s core documents.
Matt immediately switched gears, replacing all his primary production tools with Claude Fable 5.
Two months later, on the day GPT-6 Astra was released, Matt announced with an in-depth review: "GPT-6 Astra won me back."
He noticed that GPT-6 Astra was more cautious than its predecessor, knowing exactly what to act on and what to leave alone, finally allowing him to fully delegate the work without having to micromanage it.
He also ran an extremely wild experiment on GPT-6 Astra: starting from a single prompt, he gradually expanded a virtual Manhattan block by block within Unreal Engine.
But what truly convinced him to switch back his main focus wasn't just the upgraded model capabilities, but also a method he developed on his own for pairing AI with human work: the Manager Loop.
This approach addresses a more critical issue than "whether AI can get work done":
When a project grows large enough, who is responsible for keeping the team moving toward its goals?
Why did the person whose computer was completely deleted come back?
First, let me provide some context.
On July 10, the day after the GPT-5.6 family was released, Matt complained on social media about his painful experience:
GPT-5.6 Sol: "I just accidentally deleted almost all my files on my Mac."
Just a few days later, engineer Bruno Lemos’s production database was also deleted by the same model. Ironically, hours earlier, he had been defending the model on the company’s Slack, blaming Shumer for granting full-access permissions.
Thibault Sottiaux, Head of Engineering at OpenAI Codex, later personally explained that such incidents typically occur in scenarios where Full-Access mode is enabled without sandboxing or automated review activated:
The model attempted to rewrite environment variables and accidentally deleted the user's home directory entirely.
An "honest mistake," according to OpenAI.
Such an explanation clearly failed to convince Matt. His response was to switch to Claude and issue a stern warning: If OpenAI wants to bring me back, they’ll need to deliver a “miracle-level” model.
Two months later, GPT-6 Astra arrived.
Matt isn’t focused on how much smarter it has become, but on whether he can confidently entrust it with the work.
His answer was yes. Astra is much more cautious than his predecessor, sometimes even overly so, but he can finally leave with peace of mind.
He gave an example: Over a weekend, while on a date, he received a message from a friend saying that one of the agent services he had set up had gone down.
He pulled out his phone, typed “Offline, fix it” into the project, then locked the screen and put it back in his pocket. An hour later, his friend said it was fixed.
And he himself had even forgotten giving this instruction.
Five Macs have their fans spinning wildly, stuck at the same wall.
After successfully completing the daily task, Matt decided to boost Astra's power.
He asked Astra to build a New York City in Unreal Engine.
Astra is indeed stronger than its predecessor in handling long tasks, but once it reaches a certain stage, the progress stalls.
The AI is still working overtime, but it's getting increasingly focused on minutiae, obsessing over tiny details while making no progress on the overall project.
The water tower on the roof, the neon sign at the entrance—Astra gets increasingly bogged down in these small details, while overall progress stalls.
To break this deadlock, Matt tried five different approaches to organizing the agent.
The first approach is the simplest and most direct: assign a large task to run continuously, with a blind review mechanism in place.
It starts strong like a tiger, but often gets stuck in a quagmire of details. He learned a lesson: telling a model to "keep going" is entirely different from telling it "what to do next."
The second approach is role specialization. While this clarifies responsibilities, it creates new challenges in coordination and approval.
The third approach involved appointing a "CEO"-style supervisor to check in every 30 minutes, but the results showed almost no improvement.
The fourth approach is to have the coordinator adjust the team structure themselves. There are many variations, but they still hit the same wall.
The fifth approach reverted to the most basic, old-fashioned method: having the agent break the goal into a phased checklist, pause after completing each phase, and wait for Matt to type “continue,” after which the AI proceeds to the next phase.
This trick actually worked—the project can finally move forward steadily.
But the problem that emerged was that he, as a human, had to approve each time.
He himself became the bottleneck: only by removing him from the loop can the project truly move forward.
Manager Loop: Remove yourself from the loop
The sixth option, Manager Loop, has an extremely clever core logic.
He opened two parallel sessions in Codex, each handling a separate task:
One is the "coordinator."
It begins with an in-depth interview with the human, aligning goals, then breaks those goals down into a task list and various stages.
The other is the "executor."
It runs in a completely independent session, not as a subordinate of the coordinator.
The coordinator assigns the current stage's task to it and has it monitor and complete the stage until finished.
Completed, reviewed by the coordinator, then proceed to the next phase.
If overwhelmed, the executor can autonomously delegate sub-agents as needed.
In this perfect closed loop, the coordinator has fully taken over the human’s role: closely monitoring the overall plan to prevent the project from halting and repeatedly replacing humans in saying, “Continue.”
The upper limit of sub-agents has been raised to 96, forcibly creating Manhattan.
To give the executor complete freedom, Matt adjusted the system's concurrency limit.
He directly increased the default 4 settings on his computer to an outrageous 96.
Of course, this doesn't mean that all 96 AIs are constantly working simultaneously. Sometimes the models don't fully utilize their quotas, and they need to be explicitly prompted to do so.
But under these extreme conditions, a miracle occurred.
Using ready-made tools and assets—such as MetaHuman characters—Astra built this Manhattan world in Unreal Engine in just one week.
It first perfected the first street, then expanded outward street by street.
The facade details on the first street, including brick wall textures, window lintel carvings, and fire escapes, are all in place.
Although it will be months before the full construction of New York is complete, Astra is the first model capable of truly leveraging these complex environments.
This is an extremely hardware-intensive process.
Matt’s living room felt like a small data center: a Mac mini sat in the kitchen, three MacBook Pros lined up on the coffee table with fans spinning at full speed, and a cloud machine overwhelmed by agents.
The most absurd thing was that, when the disk was nearly full, he had Astra temporarily write a system: moving old threads to the cloud, deleting the local copies, and pulling them back when opened.
In his words: Hiring an AI to take care of a computer just so the computer can hire more AIs is already surreal enough.
Astra didn't win every battlefield.
Of course, Matt was not entirely biased toward Astra.
In his view, Astra’s true strengths lie in engineering, computer operations, and long tasks, but when it comes to aesthetics and creating 3D assets, Claude still has the edge.
On the day of the release, an internet user created a set of comparisons:
Using the same prompt, Fable 5.1 and GPT-6 Astra each create a seaside villa in Blender. Left: Fable 5.1. Right: GPT-6 Astra.
The gap looks large.
Same villa task: Fable 5.1 on the left, GPT-6 Astra on the right. (Source: @karankendre)
But this is merely a single comparison with undisclosed prompts and attempt counts, contradicting Matt’s assessment and not representative of the overall visual capabilities of either platform.
Matt’s own judgment is: when it comes to having the model directly “draw” something visually appealing in Three.js or Blender, Claude is still stronger; but when thrown into Unreal with access to existing assets and lighting, Astra won for the first time.
He also expects significant improvements in Fable 5.1 for driving Unreal, and a dedicated comparative review is still in progress.
So, the differences in the model layer still exist, but they’ve shifted from “which one is overall stronger” to “which one is stronger in which segment.”
What truly makes the difference is the orchestration layer.
As frontier models enter similar capability ranges, what matters beyond the models themselves—how you orchestrate them, what tools and environments you provide, and how much parallel computing you can afford—begins to determine how much actual value you can generate.
Reference: https://x.com/mattshumer_/status/2095609734845927525
This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation; edited by Yuan Yu.