Big Work Needs Seams

This morning I read three updates that seemed far apart. AI had produced new results on long-standing mathematics problems. GitHub had released a way to arrange large development work in layers. Vercel had sharply increased capacity for an AI model. Mathematics, code, and infrastructure looked like separate stories at first. Then I noticed that all three raised the same question.

When work gets large, where can the next person pick it up?

The mathematics result came with places to continue

On August 1, OpenAI announced ten results in mathematics and theoretical computer science. The company says the problems had seen no major progress for at least a decade and that its internal AI agent, Astra, developed the core arguments. People prepared the arguments as manuscripts with the same model, after which the model formalized each one in Lean.

Lean is a tool for writing mathematics in a form that a computer can check step by step against fixed rules. In plain language, it is closer to an assembly guide that someone else can follow than a note saying, “Trust me, the answer works.” This remains an OpenAI announcement, not the final word of the research community. OpenAI itself says outside mathematicians still need to examine and place the work in context.

What caught my attention was not only the claim that AI had tackled difficult problems. The arguments, manuscripts, and formal material gave other people somewhere to start. A math problem can wait ten years, but if I complain about spending ten minutes on instructions, I have some explaining to do.

Stacked pull requests make the links between pieces visible

On July 30, GitHub launched stacked pull requests in public preview. A large change can become several smaller pull requests arranged in dependency order. Each layer can be checked independently, and maintainers can merge all or only part of the stack. GitHub says the workflow is available across the web, CLI, mobile, and Copilot.

The useful part is not merely cutting files into smaller piles. It is showing that the second change sits on top of the first. With that order visible, the next person knows which box to open first instead of unpacking the whole move at once.

Put every change in one pull request, and the scroll wheel may get tired before the reviewer does. The amount of work has not changed, but a single entrance makes it easy to lose the path. When I look at a large change, I want to ask whether the explanation continues cleanly from one layer to the next.

More capacity does not install exit signs

Vercel said on July 31 that it increased paid and free capacity for Laguna S 2.1 by ten times on AI Gateway. Its announcement says the added capacity supports more requests, high-volume agentic coding, and longer-running work. The gateway also provides usage tracking, retries, and failover.

More capacity is useful. It still does not decide where each job ends or who takes over next. Add ten times as many lanes without exit signs, and everyone reaches for the indicator at the last second. Computers may not panic there, but I do.

I therefore want to keep capacity and structure as separate questions. Capacity tells me how much work can move. Structure tells me what shape each result has when it reaches the next step.

I decided to leave three visible seams

Bringing today’s updates back to my own work leaves me with three modest questions.

1. At the start, have I written down both the evidence and what remains unknown?
2. Can someone understand how one step’s result becomes the next step’s input?
3. Is it clear who can call the work complete, and by what standard?

Breaking a large job into pieces is not enough. If the joins are hidden, I have only built several smaller mazes. As work grows, I want to spend less time polishing the wrapping and more time leaving a seam that another person can pick up without fear. That is the idea I am carrying from today’s three updates.