Anthropic’s IPO Eve Looked Like a Victory Lap. Then Claude Code’s Core Team Started Leaving.

On October 23, Anthony Morris, creator of Claude Code’s desktop client, announced on X that he was leaving Anthropic. The next day, core engineer Thariq Shihipar followed. On November 3, product lead Cat Wu confirmed that October 30 had been her last day. On November 4, she said she was joining OpenAI to lead product management for its core product team. In less than two weeks, the founding team of Claude Code lost at least four core members.

In the same window, Anthropic filed its S-1 with the SEC. Revenue in the first half of 2026 was $16.29 billion, fourteen times the same period a year earlier. Valuation expectations ranged from $1.5 trillion to $2 trillion. The plan was to list on Nasdaq in November. In the risk factors section, one sentence said the company might “soon” meet its Responsible Scaling Policy’s definition of “AI-driven dramatic acceleration of progress,” which is the threshold for recursive self-improvement. The next sentence said that no current model had exceeded that threshold, and that the company was committed to maintaining meaningful human oversight and control over frontier model R&D for the foreseeable future.

The two texts appeared in the same month and had no causal relationship. They shared a premise: when the speed of technological diffusion exceeds any single vendor’s ability to maintain a generational gap, even the question “who is irreplaceable” needs a new answer.

1. The role that left was product definition

Cat Wu was not an ordinary product manager. When Boris Cherny posted the early Claude Code prototype in an internal Slack, she was among the first users. She later wrote in her farewell post that she had once let Claude Code run for more than ten hours to build a reinforcement learning environment, completing the task “without writing a single line of code myself.” That experience led her to transfer into the Claude Code team on the spot. She and Boris were called the project’s “Batman and Robin” by outsiders. One handled product direction; the other handled technical development. Together they turned an internal tool that initially drew little attention into one of Anthropic’s most important revenue engines. Among enterprise users of Claude Code, the top 5% by spending had average monthly outlays about twenty times those of non-users.

Another detail from her farewell post explained the trajectory better than any resignation reason. Claude Code’s default behavior had shifted from “users manually approving most permission requests” to “trusting Claude’s judgment to auto-approve.” Claude now “executes large-scale migrations and builds new products.” When the product itself increasingly does not need to be defined, the role of the product lead shifts with it.

Anthropic’s hiring data supports that shift. Within a year, the share of junior technical roles in total hiring fell from 27% to 9%. The runtime logged by Claude agents directed by employees was six times that of human operators. Engineering manager Felix Rieseberg announced that his team was eliminating coding interviews entirely. His reason: “the impact of senior engineers is no longer proportional to hand-written coding ability.” Evaluation would focus on “problem definition.” Claude Code lead Boris Cherny was more direct. The boundaries among engineering, product, and design were merging, and traditional job labels would be replaced by five new roles based on behavioral patterns.

Thariq Shihipar worked on context engineering and dynamic workflows and once cut Claude Code’s system prompt by more than 80%. Anthony Morris worked on bringing the terminal tool to desktop and lowering the barrier to use. That work remained important. Its importance was shifting from shaping user experience that could not be replaced to building operating frameworks for autonomous systems that could be replaced. The departures did not mean people were no longer needed. They meant the type of people needed had changed. The work was moving from defining product experience to defining the Harness’s evaluation functions, optimization objectives, and operating boundaries. That is the core focus of RSI engineering.

2. Harness turned RSI into a roadmap, but fully autonomous RSI remains a distance

Equating the departure wave with “Harness has achieved fully autonomous RSI” is an overreach. Treating Harness as the most realistic engineering path to RSI has solid technical grounding.

In her systematic review of Harness Engineering, Lilian Weng argued that progress in RSI depends on more than models becoming smarter. The Harness around the model is the system layer for orchestration, tool use, context management, artifact storage, and result evaluation. Its engineering is becoming the main focus of self-improvement. The abstraction level of optimization targets is moving upward: from instruction prompts to structured context to workflows to Harness code, and finally to optimizer code itself. “Letting Harness modify Harness” has been validated by several methods: STOP, Self-Harness, Meta-Harness, AlphaEvolve, and Darwin Gödel Machine. DGM raised SWE-bench Verified scores from 20% to 50%. Within specific technical domains, the loop of system self-iteration has closed at the engineering level.

Between a closed loop and full autonomy lie several hard constraints. The same review listed seven bottlenecks for full RSI: weak and ambiguous evaluators, lifecycle management of context and memory, scarcity of negative results, diversity collapse, reward hacking, long-term success that is hard to measure, and the role of humans in the loop. One empirical result matters: STOP improves performance on GPT-4 but worsens it on GPT-3.5 and Mixtral. Recursive structures need a sufficiently strong base model. Self-improvement is not unconditional.

Anthropic’s S-1 wording matches that picture. The company said internal evaluations “and expectations for future models” suggested it might soon touch the RSI threshold. It then said: “we have not judged any current model to exceed this threshold.” It promised that “without meaningful human oversight or understanding, even if technically feasible,” it would not let Claude fully autonomously design and build later generations of frontier models. The filing did two things at once. It used the RSI narrative to support IPO valuation. It used a human-oversight commitment to set a safety boundary for that narrative. The departure wave arrived as this dual narrative took shape. It looks more like a side effect of product roles being repriced under the RSI roadmap than evidence that fully autonomous RSI has arrived.

3. Moats shorten through two paths: distillation upstream, open source downstream

If local RSI at the Harness level is turning technical capability into a reproducible output, two other forces are compressing the window in which that output becomes durable advantage.

Distillation extracts capability upstream. Jefferies analyst Edison Lee and others judged that widespread distillation is “further lowering industry entry barriers.” New players can enter the top-tier model market with tens of millions of dollars in funding. Without distillation, the gap between some challengers and frontier models could widen to 18 months. Independent training without distillation can cost 6 to 10 times as much as the distillation path. Even when closed-source models do not disclose Chain-of-Thought, multi-turn API calls and synthetic data generation can still extract implicit knowledge. Building a stronger model and turning that strength into a gap others cannot close are increasingly different tasks.

Open source accelerates diffusion downstream. After DeepSeek open-sourced its internal inference engine DeepInfer in October 2026, vLLM and SGLang completed T0-level adaptation in less than a week. In SGLang’s decoding test at 256 requests per GPU, per-step latency was 28.88 milliseconds, about 7.4% shorter than DeepInfer’s 31.19 milliseconds. Peak prefill throughput was about 4% higher. vLLM’s new-token throughput in the prefill stage rose by 35.6% to 43.9%. Under GB300 rack deployment, per-GPU throughput reached 7.9 times the previous single-node result. DeepInfer was initially built on vLLM and drew on SGLang’s design. After DeepSeek made its polished implementation public, the two teams absorbed the results and kept optimizing. The loop ran from community to community and back again.

DeepSeek’s strategy is not to make others unable to catch up. It is to make the whole industry run faster while requiring itself to keep a faster pace of innovation. When leadership in inference engines is caught or surpassed in units of days, the shelf life of any single vendor’s engineering-efficiency moat becomes very short.

4. The last moat sits in the rules, and the rules are losing legitimacy

When technological diffusion outruns any single vendor’s ability to hold a generational gap, the center of competition moves. The question changes from “who builds it first” to “under what rules.”

Recent moves by Anthropic and OpenAI point there. Dario Amodei published an open letter calling to “slow down the frontier.” He cited RSI risk as the core reason and said AI’s “ability to build the next generation of AI” is “beginning to happen across the industry, including at Anthropic.” In September, OpenAI proposed international technical standards for RSI. It defined RSI as the process by which “the proportion of work required to build successive generations of AI performed by AI systems continues to increase.” If first movers shape the threshold definition, evaluation methods, and oversight requirements, later entrants may replicate the technology but pay extra in compliance costs and time. The first mover’s advantage is not a permanent monopoly. It is making rules favorable to itself the industry default before diffusion is complete.

That logic has a weak point. One reading of the coordinated call by Anthropic, OpenAI, and Musk to “slow down AI” is regulatory capture: using safety as a name to erect barriers and put open-source models and challengers at a structural disadvantage. The timing invites the reading. Anthropic began warning loudly about RSI as rumors circulated that Google DeepMind might already be close to the capability. A global slowdown would require China’s cooperation, which is nearly impossible in the short term. A government-led moat becomes the alternative. The same comparison points to earlier scares: quantum computing breaking encryption and Y2K destroying software. Neither stopped technology from developing.

Rule-making power as the last moat is not a stable anchor. It depends on one premise: the credibility of the rules is not eroded by the competitive motives of the people writing them. When the most active promoters of the rules are the parties most likely to benefit, the rules lose legitimacy.

Return to the timeline from October 23 to November 3. In the last paragraph of her farewell post, Cat Wu wrote about what she learned at Anthropic. She did not mention why she left. She did not mention OpenAI. Thariq Shihipar’s post was shorter. He said only that “it’s time for the next thing.” In the S-1, RSI appeared in the risk factors section, with words like “may” and “expect.” The three texts have no obligation to explain one another. They appeared together in a month headlined by “14x revenue growth” and a “$2 trillion valuation.” Technical moats are being compressed by distillation and open source in units of weeks or days. Rule moats are harder to replicate and harder to trust. First movers still matter. Their advantage depends less on “I built it first” and more on “I can make the world accept the rules I define before others build it.” Anthropic is betting on two fronts with very different time scales.