People Decide What, Agents Decide How
Anthropic just filed the receipt for Compression Theory.
On June 16, Anthropic published a study of about 400,000 Claude Code sessions from roughly 235,000 people, spanning October 2025 through April 2026. The title is "Agentic coding and persistent returns to expertise." The finding underneath it is the thesis this channel has been building since day one.
Here is the sentence that matters: people decide what to build, and the agent decides how to build it.
They measured it. In a typical session, the human makes about 70 percent of the planning decisions. The agent makes about 80 percent of the execution decisions. Planning is what to do, which approach to take, what counts as done. Execution is which files to touch, what code to write, which commands to run. The split is clean, and it holds across the whole dataset.
That is Compression Theory with a control group.
The strong signal: value migrates up
Every technology wave compresses one layer of work. When a layer compresses, the work doesn't vanish. The value migrates to the adjacent layer that hasn't been compressed yet. Spreadsheets compressed arithmetic, and the value moved up to modeling and judgment. GPS compressed navigation, and the value moved up to the apps built on top of it: ride-hailing, delivery, real-time logistics.
AI is compressing the creation of software. This report is a direct read on that happening live.
Watch the "how" layer collapse. The share of sessions spent fixing broken code fell from 33 percent to 19 percent in seven months. Debugging, the most implementation-bound task there is, nearly halved. Writing the code, wiring the functions, chasing the syntax: that work is folding into the agent.
Now watch the value move. The work around the code grew to fill the space. Operating software went from 14 to 21 percent of sessions. Writing and data analysis roughly doubled, from about 10 to 20 percent. The value of the average task rose about 25 percent over the window, with build-type work up around 43 percent.
The "how" got cheaper. The work moved up. That's a perfect fit for the argument we've been making with Compression Theory, and now the data confirms it.
The payload: domain beats coding
Here is the part to tattoo somewhere. The thing that predicts whether a session succeeds is not whether the person can code. It's whether they understand the problem.
Anthropic rates expertise as task-specific, not by job title. Their own example: an accountant who has never written a line of Python, but who tells the agent exactly which reconciliation rules to enforce and catches the edge case it mishandles at month-end close, is an expert at that task. And any of us, however much we've shipped, is a beginner the moment we're working a problem we don't actually understand yet.
The success numbers track it. A novice-rated session reaches verified success 15 percent of the time. Intermediate or better lands at 28 to 33 percent. When a session runs into trouble, novices abandon it 19 percent of the time. Everyone else gives up at 5 to 7 percent. The gap is command of the domain, not command of the keyboard.
Then there's the line that ends the "but you still have to be a developer" argument. In sessions that produce code, every one of the ten largest occupations lands within seven points of software engineers on success. Management scored slightly higher than the engineers.
Read that twice. Lawyers, analysts, scientists, managers: directing the agent, shipping working code, succeeding at nearly the developer's rate. Anthropic's own phrasing is that coding agents are making a coding background less relevant to successful programming, while rewarding people who understand the problems they solve.
One honest boundary
This study is about interactive coding sessions with a human in the loop. It explicitly excludes headless and pipeline usage. It is not a study of agents shipped into products at runtime, and it makes no claim about how a business should deploy them. It corroborates where value migrates. It does not settle how you wire an agent into a shipping product. That's a separate argument for a separate post.
I've said this before and it bears repeating: shipping software is shipping trust. Code that you don't understand is not a great recipe for trust. IMO we're either missing a critical validation capability (particularly for agents, remember "promises"?), or we still have developers who can call BS on bad decisions. The "how" is compressing, but the story is far from over.
But on the core claim, the results are compelling. The compression happened. The value moved up. The people who win are the ones who understand the problem.
The skill that survives is the product-management skill. Define the problem. Validate that it's real. Check the constraints. Decide what's worth building before anyone builds it. That used to be one job title. The data says it's becoming the job, across every field that touches an agent.
We are all product managers now, and it looks like Anthropic agrees.
Source: Zoe Hitzig, Maxim Massenkoff, Eva Lyubich, Ryan Heller, and Peter McCrory, "Agentic coding and persistent returns to expertise," Anthropic, June 16, 2026.