The Agent SDK, The S-1, And The Slop Problem

00:00:04 Anthropic filed its S-1 today. The filing puts first-half 2026 revenue at $16.3 billion. More than 11,900 customers are generating at least $100,000 annually, up from 950 a year earlier. About 1,900 of those have crossed $1 million, and roughly 250 have hit $10 million.

00:00:24 Gross margin came in at 85.4 percent for the first half, and the second quarter alone produced $974 million in operating profit under standard GAAP accounting. Nate B Jones highlighted one number that keeps coming back to mind: among publicly traded Fortune 500 customers, spending with Anthropic averages less than one dollar for every thousand dollars of total expenses.

00:00:52 The intelligence-as-input angle is familiar — companies buy Claude the way they buy cloud compute, the way they buy legal counsel, as an input that makes existing people and capital more productive. The S-1 shows just how early that adoption still is, even at a company this size.

00:01:13 The 11,900 substantial customers include all ten of the largest banks and nine of the ten largest pharmaceutical companies. That's a mature customer list on paper, but at $100,000 a year, that can still be a small deployment in a small part of the business. Even Anthropic's ten largest Fortune 500 customers, who spend on average more than eight times the qualifying group's average share of expenses, are still below one percent of total expenses.

00:01:46 The S-1 describes a progression that starts with employees using Claude for their own work, someone using it for code and analysis, a useful application gets discovered, and usage spreads. More than half of substantial customers now use both the Claude platform and at least one of Anthropic's own applications like Claude Code — up from about 30 percent in the last quarter of 2025.

00:02:15 The expansion path goes from engineering to other departments, then into processes, and finally into products. The filing shows this happening. Co-Work, the non-technical product, is growing faster than Claude Code did. Novo Nordisk's NovoScribe project started with clinical study reports and has extended into device protocols and patient materials.

00:02:41 They're now testing Claude Science in selected R&D workflows. Eighty percent of first-half revenue was consumption-based. Customers pay according to usage, and their activity continues to affect the relationship after the purchase. The September quarter saw about 80 percent of revenue from model versions released during the preceding six months.

00:03:07 New capabilities are doing the heavy lifting. Opus 5.5 reduced customers' cost per task by about 40 percent compared with Opus 5, through a combination of lower pricing and improved task efficiency. The paradoxical thing about making a resource cheaper is that you tend to expand the uses to fill the space — what Jevons called the efficiency rebound effect.

00:03:34 Anthropic's data supports this: they've generally seen customers expand usage after releases that combine greater capability with lower task costs. The business question is whether customers buy enough additional work to outweigh the lower cost of each task. That's something investors will pour over in the months to come.

00:03:58 Right now, what stands out is that these companies are still learning how to use this product, and the product is gradually becoming useful for more and more things. The S-1 also shows Anthropic moving from tools to drugs. They acquired Mithril Medicine, a one-person company with two rare muscular-dystrophy programs, and are continuing those programs internally.

00:04:25 There's a tension there that the filing doesn't fully resolve: a pharmaceutical company has to assess a supplier's possible long-term role in developing medicines, and a software company faces a more immediate product decision about how much of what it sells could become a standard feature of the model supplier's own application.

00:04:50 Better Claude models can make a supplier's product better while bringing Anthropic closer to being able to offer similar work itself. Anthropic is supplying intelligence to drug developers while also becoming a drug developer itself. The commercial data policy says commercial inputs and outputs aren't used for model training by default, which protects confidential data.

00:05:18 But trust, control, and competitive overlap can affect adoption alongside capability and price. The S-1 reads less like a financial document and more like a window into how one of the most successful companies in history views the problem of scaling AI inside organizations.

00:05:39 The filing doesn't get into workflow complexity — how quickly Anthropic can scale across departments, how sticky a given deployment is, what happens when the next deployment builds on the last — and those are the real questions. No one has mapped workflow stickiness in corporate America yet, and companies aren't going to share that secret sauce.

00:06:05 But Anthropic is effectively betting that training 10,000 engineers through the Claude Frontier Academy will give them a jump start at getting through that bottleneck.

00:06:17 Cursor released the Agent SDK today. It's the framework they use to build their own managed agents — Bugbot, Security Reviewer, Rollouts. The goal is to let teams define agents as code and deploy them with a single command. Each agent is organized like a small software project.

00:06:38 Instructions sit alongside code for tools, connections to outside services, and evals used to test them. Prompts are written in Markdown. Tools are defined as TypeScript functions. Everything goes through version control and review, just like the rest of the software.

00:06:58 You can deploy to Cursor infrastructure with agent-sdk deploy, or run the agent on your own infrastructure — each one runs as a single Node process that can be packaged in a container. The Agent SDK supports evals and event replay out of the box. A team within Cursor built a kb-reviewer agent in about a day that runs on every pull request in their main codebase, flags changes that contradict recorded lessons from their knowledge base, and encourages engineers to reconcile discrepancies.

00:07:34 The PGA Tour uses the Agent SDK to power their entire software development factory. The platform play is the real shift here. Until now, building an agent generally meant adopting a product designed around someone else's workflow or starting from scratch. The Agent SDK lets you build around your own workflow, respond to events your team handles, and follow policies your team can inspect and change.

00:08:04 Cursor is building around its editing tool, but teams can build agents on Cursor's infrastructure, integrate with GitHub, GitLab, or Bitbucket, and follow their own policies. It's a different angle from the vertical AI tool companies. You're not buying a product that does code review; you're getting the framework to build the code-review agent that fits your team.

00:08:31 The TypeScript tool definitions and Markdown prompts are the right choice here. They keep the agent definition in the same language and format that engineers already use. If you're going to put agents in your version control, they need to read like code. A practical detail: agents can hand a task to another specialized agent with its own tools and sessions.

00:08:58 They can divide a job among subagents working in parallel. That composability is what makes the SDK interesting beyond the template layer. The filing mentions evals as a first-class concern — rapid development with assurances that critical workflows execute the way you expect.

00:09:18 If the evals are good, you can deploy agents with confidence. If they aren't, you've got another problem to solve. We'll see what happens when developers start shipping agents built on the SDK.

00:09:33 Valve spoke at Steam Indie Connect in Korea on Tuesday, and the session covered something that's been building for a while: the AI clone problem on Steam. Tom Giardino and Marc Wilkinson from Valve's Steam Business Team opened with a straightforward update. Steam reached 43.5 million concurrent users at peak.

00:09:55 Total Steam revenue in South Korea grew 126 percent over the past five years. Premium game and DLC revenue increased by 70 percent. In-game purchase revenue surged by 388 percent. The AI cloning threat started with something specific: an indie developer who received an extortion email saying, 'If you don't pay, we'll clone your game with AI and release it first.' Giardino expressed regret — this was the first time he'd heard of such cases.

00:10:26 He noted that clone games existed long before AI. Spotting an interesting concept during Next Fest and quickly making a similar game has always happened; AI has simply accelerated that pace. Valve's response was characteristically blunt. They don't want to be gatekeepers telling developers which tools they should or shouldn't use.

00:10:50 Whether to use AI in development is a decision for the developers themselves. Rather than trying to filter or block, they're leaning on what they've always said: make the genuine version the best version. Launch isn't the finish line — it's the starting line, Giardino told the audience.

00:11:10 The features, personality, and creativity that stem from your perspective cannot be easily copied the way a single mechanic can. Valve has historically not been aggressive with DRM or anti-piracy measures. They've always preferred making the legitimate version the best available option.

00:11:31 The AI slop issue came up through Next Fest, the event where new games get exposure to players. Marc Wilkinson explained the visibility structure: for the first two days, games are shown on a virtually random basis to prevent system abuse and give every game fair exposure.

00:11:50 After that, personalized recommendations kick in. The complaint about low-quality games flooding the event came from those first two days — when the random phase exposes everything, including the slop. We recognize that the trade-off is some low-quality games get exposed, Wilkinson said.

00:12:11 We are fully aware of the situation, and a significant number of people are thinking about how to solve this problem. There's a parallel to Valve's philosophy on greenlighting and self-publishing. When Steam Greenlight launched around 14 years ago, developers asked if Greenlight games could be filtered out.

00:12:33 When Steam Direct was introduced, requests came in for a checkbox to hide Steam Direct games. Every time, Valve chose to improve the shopping experience rather than turn developers away or try to pre-filter good games from bad ones. Giardino said they don't think they're perfect and there's plenty of room for improvement, but making Steam a smarter and more dynamic platform is the starting point of their philosophy.

00:13:03 The extortion question is closer to an issue for law enforcement than a platform problem, Giardino noted. But Valve said they'll look into reported cases and help however they can. There's a distinction between slop as a quality problem and slop as a market-structure problem.

00:13:23 When anyone can generate a game with AI tools, the barrier to publication drops dramatically. Valve's approach of improving discovery and recommendation rather than restricting access is a design philosophy, not just a practical choice. It's also a bet that community and continuous improvement — the things that take real human effort — are more valuable defenses than any technical barrier.

00:13:51 The clone games aren't the problem — the problem is the incentive structure that makes them a good use of someone's time. If your game is good enough that someone can make money by cloning it, you've probably already won. A Reddit post from the ClaudeAI subreddit this morning shows the other side of the coin.

00:14:13 A project manager there made a pull request with Claude — generated the code, had Claude review it, and now the team has to do the human review on top. Standard procedure for AI-created PRs in their workflow, the author noted. The post blew up fast enough that the author couldn't follow all the replies.

00:14:35 On one end, Valve is dealing with AI-generated games flooding a platform. On the other, a project manager is using Claude to generate actual production code. The tool is the same. The outcome depends on what you're building and who's accountable for it.