Agentic Code and the Infinite Game: What Claude Code Means for Research and Innovation

On New Year's Day, programmer Steve Yegge launched Gas Town, an open-source platform that lets users orchestrate swarms of Claude Code agents at once to assemble software at blistering speed. The results were both remarkable and overwhelming. As one early user put it, there was "really too much going on for you to reasonably comprehend." That sentence captures where we are: agentic code has gone from a demo to a powerful force, fast enough that in June The Economist called Anthropic's commercial success "astonishing," warning that it makes the company a target. Five months later, Anthropic went public in the largest global IPO on record, priced at $83 per share.

But the more intriguing question is not how fast the code gets written. It is what happens to the humans doing the research, and to the game they think they are playing.

What agentic code does for research

Claude Code is not a chatbot that answers questions about code. It is an agent: it writes code, runs commands, and works through files on its own, pursuing a goal across many steps. In software, that means an engineer can hand an agent a task and review the result. In both research and medical imaging, the same shift is underway: agents that build analysis pipelines, segment scans, run experiments overnight, and keep a reproducible record of everything they tried. The promise is the one discussed in the last post: the AI stops being a tool that follows instructions and becomes a collaborator that figures out the instructions itself.

The numbers explain why the market noticed. Claude Code reached general availability in May 2025 and by February 2026 was running at a $2.5 billion annualized revenue rate. Anthropic's revenue went from zero when Claude launched in early 2023 to $17 billion in a single quarter, with nearly 12,000 business customers spending $100,000 or more a year. At its November IPO, the first trade printed at $111.50, up 34% from the $83 offer price.

The Economist's chart tells the deeper story: Anthropic's annual recurring revenue has raced past OpenAI's. Four-fifths of Anthropic's revenue comes from business customers, against about 40% at OpenAI. The coding agent, capable of tasks that would take human coders hours or days and producing work that is easy to check, turned out to be the durable business. In May, the firm was valued at $965 billion, ahead of OpenAI's $852 billion in March, and it told investors it would post an operating profit in the second quarter, the first AI lab to do so.

The hidden cost: AI brain fry

Speed, however, has a human price, and it is now measured. A study published in Harvard Business Review (July–August 2026) surveyed 1,488 full-time workers. It gave the phenomenon a name: "AI brain fry," mental fatigue from overseeing AI tools beyond one's cognitive capacity. Fourteen percent of AI users reported experiencing it.

The study found that the most taxing form of AI work is not using the tools but overseeing them. Workers whose AI work required heavy oversight expended 14% more mental effort and reported 12% more mental fatigue. Productivity rose as people went from one AI tool to two or three used at once — and then it dipped. Those experiencing brain fry reported 33% more decision fatigue and 39% more major errors, and their intent to leave rose from about one-quarter to about one-third.

That is the Gas Town problem in a lab coat. An imaging researcher can now set a dozen agents running: one segmenting scans, one testing a model variant, one drafting the methods section, and end the day exhausted not from the research but from managing the research. As one engineer quoted in the study put it, "I end each day exhausted, not from the work itself, but from the managing of the work."

The same study offers a counterpoint, and it matters: when AI replaced routine, repetitive tasks (the toil), burnout scores fell 15%. The tools are not the problem. The oversight load is. And the checkability that made Claude Code a business success points to the fix: agents whose work is easy to verify require far less oversight than agents whose work is not.

Research is an infinite game.

Simon Sinek's book, "The Infinite Game" (2019), provides an insightful perspective on business and research. Sinek argues that these fields should not be viewed as finite games with clear winners and losers that culminate in a final countdown. Instead, they resemble an infinite game, where the goal is not to win but to continue playing and advancing a greater cause. He outlines five essential practices for playing this game effectively:

1. Advance a just cause.

2. Build trusting teams.

3. Study worthy rivals.

4. Prepare for existential flexibility.

5. Demonstrate the courage to lead.

Anthropic is, knowingly or not, a case study in all five. Its just cause is written into its structure: it is a public benefit corporation, legally bound to weigh the public good against commercial success. Its worthy rival is OpenAI and not an enemy to crush but a mirror that reveals where Claude must improve. Its existential flex was Claude Code itself: a company built on chatbots bet its future on a coding agent, and that bet became its growth engine. And its courage to lead showed in going public while keeping Claude free of advertising, in a market that rewards the opposite.

The Economist's briefing is, in Sinek's terms, a story about what happens when a finite player meets an infinite one. The briefing frames Anthropic as almost monastic, featuring a headquarters that embodies this atmosphere. A chief executive writes encyclicals on AI safety while facing a government determined to take control. This has led to the firm being labeled a "supply-chain risk," which requires an export license for every sale to foreign entities. Additionally, the government has mandated that Anthropic cut off non-American individuals, including its own staff, from accessing its latest models. David Sacks warned that the rapid growth of Anthropic could lead to "unprecedented control over the most important technology of our time," comparing it to Rockefeller's oil empire. Instead of fighting for dominance in the industry, Amodei advocates for changing the rules of the game. He proposes regulations similar to those of the Federal Aviation Administration, requiring frontier models to undergo standardized technical tests and audits. This approach would allow all companies to continue competing in the market. As Dean Ball, a former Trump administration official, points out, there is a risk of achieving government "dominance over AI" instead of fostering American "dominance in AI."

Research works the same way. The just cause in medical imaging is not publishing first or beating a rival lab; it is better, earlier, more reliable diagnosis for patients. A rival lab is worthy when it reveals what your own methods miss. And existential flexibility is a lab willing to rebuild its pipeline around agents while redesigning the work so humans oversee two or three agents well instead of drowning in twelve.

Bottom line: Agentic code will keep getting faster, and the companies and labs that win this quarter's finite game will be the ones that ship the most. But the infinite game belongs to those who protect the scarcest resource in the room — human attention — and keep playing for the cause that outlasts any single model, paper, or IPO. The agents can write the code. The judgment about what is worth doing, and why, is still ours, and it is the part of the work that makes the rest matter.