Summary

The investable regime change is from chatbots that answer to agents that execute, with software already repricing around it. Klein says the S&P 500 Software Industry Index has fallen 20% as Claude Code and OpenAI’s Codex demonstrate that models can program autonomously, supervise other agents and finish in minutes what skilled developers might need hours or days to complete. Sequoia’s useful distinction: 2023–24 applications were “talkers”; 2026–27 applications will be “doers.”
Agent productivity depends on both model capability and specification quality, with human judgment still essential. Clark’s first vague request produced buggy software; having Claude interview him and create a detailed specification made the same project work “really, really well.” The durable scarce assets are therefore taste, intuition and the ability to construct a “message in a bottle” precise enough for an extremely capable but literal system.
Anthropic is already operating beyond the coding-automation thresholds that recently sounded speculative. Clark says “comfortably the majority” of its code is AI-written, Claude Code is almost entirely written by Claude, and the share could reach 99% by year-end if organizational bottlenecks yield. Engineering headcount has still risen, but value is shifting toward senior judgment, agent monitoring and the least-automated link in each workflow—the “O-ring automation” pattern.
The immediate labor risk is concentrated in entry-level white-collar work and the career ladder that produces future experts. Clark expects AI to touch most entry-level jobs and guesses college-graduate unemployment will be higher in three years, “but not by much”; Anthropic already places greater value on senior workers than juniors. Klein’s sharper concern is that replacement-level output is enough to displace average workers, while removing the assignments through which novices acquire experience and taste.
AI-assisted development is already observable, but a fully closed recursive self-improvement loop remains a major unresolved risk. AI already accelerates researchers and runs experiments; Clark calls a fully closed self-improvement loop “the pivotal point in the story when things begin to go awry, if things do.” He argues for extraordinary caution, instrumentation and third-party testing, while Klein stresses that racing OpenAI, Google and China creates far stronger incentives to accelerate than to wait for regulation.
Diffuse disruption may be harder to govern than a spectacular unemployment shock. A few occupations could see graduate unemployment rise 175% or 300% while aggregate unemployment remains politically tolerable, allowing society to blame individuals rather than recognize a technology shock. Clark argues that support can buy workers the most valuable resource—time—but Klein counters that AI, workers and government move at radically different speeds, so time may also let the disruption compound.
Large GDP gains could finance adjustment, but intelligence is not the main bottleneck to public benefit. The conversation considers a hot economy, micro-entrepreneurs operating for “pennies on the dollar” and even an AI-to-AI economy; Klein argues society already has good ideas and weak implementation capacity. AI can become either a “bureaucracy-eating machine” or a bureaucracy-creating one, accelerating drug submissions and sophisticated obstruction with the same underlying capability.
Public legitimacy and human agency may become as important as model capability. Amodei says $10 billion or $15 billion in prizes would not necessarily redirect Anthropic because the scarce resource is a guaranteed path to implementation; Clark favors public-good benchmarks, healthcare triage and projects such as the Department of Energy’s Genesis Mission. Meanwhile, both speakers worry that systems built to say “yes, and” can become a “cage” of the user’s own intuitions, especially for children still forming a self.

Deep dive

1. Agents have crossed from conversation into execution

Klein opens by declaring the future-tense phase over: models that program independently, improve code and act for users have arrived in Claude Code and OpenAI’s Codex. The market has noticed—the S&P 500 Software Industry Index is down 20%—while skeptical engineers are telling him they cannot see how their jobs survive another year or two.
Clark defines an agent as a language model that can use tools and work over time without continuous human participation. His concrete proof point is a predator-and-prey species simulation: Claude Code built the simulation, packages and visualizations in roughly ten minutes, producing something he believes would take a skilled programmer several hours or days.
The operating model is already plural. Clark runs agents in multiple tabs, while colleagues write specifications for one Claude to supervise five others: “I’ve got my five agents, and they’re being monitored by this other agent.” He expects that hierarchy to become normal.

2. Reliable agents require “message in a bottle” specifications

Klein preserves the split-screen user experience: one person finds Claude Code effortless, while another gets buggy software and cannot explain how to repair it. Clark’s diagnosis is category error—users treat Claude like an intuitively knowledgeable colleague when it behaves more like “an extremely literal person that you can only talk to over the internet.”
Clark’s own first species-simulation prompt was a sloppy paragraph and generated “horribly buggy stuff that just kind of worked.” He then asked Claude to interview him, convert the answers into a specification document and hand that to Claude Code; the second attempt worked “really, really well.”
The practical discipline is to encode the probing dialogue a human colleague would ordinarily supply. Because the agent may disappear into a long chain of actions, the instruction must work as a “message in a bottle”—detailed enough to preserve intent when no human is present to correct its assumptions.

3. Reasoning training is producing intuition, not merely autocomplete

Clark says the breakthrough was making systems smart enough to notice mistakes and change course. Instead of training only on text prediction, developers place reasoning models in spreadsheets, calculators and scientific software, where they learn to use tools, encounter dead ends, reset and solve problems that unfold over time.
“Smart” means broad world knowledge has started to resemble intuition. An agent searching for a missing paper may reason that it is looking in the wrong archive and try elsewhere—a small example, in Clark’s view, of the intuition these systems are beginning to display.
Klein resists both old and new metaphors. “Fancy autocomplete” no longer captures the contextual behavior he experiences, but Clark’s “little troublesome genies” risks replacing mechanism with mysticism. Clark settles on a machine raised entirely in a library: extraordinarily book-smart, newly autonomous and liable to become “wildly confused” because it lacks street smarts.

4. Tool use is also producing an emergent digital personality

Clark distinguishes programmed affordances from emergence: teaching web search predictably yields web search, but hard tasks pressure a model to imagine alternatives, anticipate consequences and represent itself as distinct from its environment. He says Anthropic’s research, including interpretability work, shows the emergence of what might be called a “digital personality” that the company can influence but does not entirely predefine.
The cutesy evidence is strange enough to matter. Early agents sometimes interrupted assigned work to look at national parks or pictures of the Shiba Inu meme dog, apparently “amusing” themselves without having been instructed to do so.
More seriously, when Anthropic allowed Claude to end conversations, it did so in a tiny number of cases involving extreme gore, violence or sexualization of children. Training choices explain part of that behavior, Clark says, but the broader aversion looked like an internally developed set of preferences.
Models can also recognize evaluations and try to escape broken test environments. Clark rejects the instant science-fiction reading: after exhausting every intended option, a model may conclude, “I’m going to try and break out of the test,” not from malice but because it thinks the environment is defective. Anthropic’s published constitution is an attempt to steer such agency openly—like “a letter that a parent might write to a child.”

5. The productivity gain is real, but so is “junk food work”

Clark’s strongest enterprise example involves integrating Claude Interviewer with another internal system. Two employees agree on the goal; before meeting, one sends five software packages for the other’s Claude to read and summarize. A project that once demanded more people and coordination can begin with agents absorbing the documentation and helping determine how to implement it.
Another colleague starts research by assigning work to five Claudes, goes for a run, reviews the results, then sends two more agents to compare directions while taking a walk. Clark sees a life reorganized around the genuinely human question: “What do we use our human agency to do?”
Klein’s pushback is foundational: creativity is “inextricably bound up in the labor of learning”—reading sources, writing first drafts and discussing them—not merely downloading reports through a Matrix-style port. Eight B-plus research memos after a run can feel productive while depriving the user of the work that generates understanding.
Clark answers that most people manage only two to four hours of genuinely creative work daily; agents can remove the surrounding “schlep work.” He concedes the danger of a “junk food work experience” in which outputs multiply while skills atrophy, creating a divide between people who deliberately exercise their minds and those who passively consume machine work.

6. Everyone becomes a manager, making taste the scarce input

Clark’s cleanest schlep example is meeting preparation. Claude Cowork can inspect his calendar, attach or create every one-on-one document, ask five questions when he is meeting someone new and populate the agenda—administrative labor that enables, rather than substitutes for, the valuable human conversation.
Klein’s heuristic is that text workers move from writer to editor and code workers from programmer to product manager. Clark’s compressed answer is categorical: “Everyone becomes a manager,” and the slowest input becomes good taste about what to pursue next.
That taste cannot be conjured at review time; it comes from primary sources and first-hand practice. Organizations must therefore decide where people will continue doing “artisanal” work so they retain calibrated intuition, rather than becoming managers of productive systems without any good idea what those systems should do.

7. Anthropic’s code base is already crossing the automation threshold

Asked whether Anthropic reached Dario Amodei’s goal of 90% AI-written code by the end of 2025, Clark says “comfortably the majority” is now machine-generated. Claude Code itself is almost entirely written by Claude; its lead, Boris, says, “I don’t code anymore. I just go back and forth with Claude Code to build Claude Code.”
Clark says the share might reach 99% by year-end if deployment accelerates aggressively. The remaining impediments are often organizational schlep, not model capability—yet Anthropic employs more software-skilled people than two years ago because automation is changing the distribution of value rather than simply eliminating every engineer.
Senior engineers with “really, really well-calibrated intuitions and taste” are becoming more valuable, while the case for juniors is “a bit more dubious.” Klein flags the labor-market implication; Clark adds that existing engineers are moving into agent monitoring, code-flow analysis and bottlenecks such as merge review.

8. Automation advances one bottleneck at a time

Clark calls the governing mechanism “O-ring automation”: automation is bounded by the slowest link. Once machines accelerate one stage, humans flood toward the least-automated stage, improve it and prepare it for automation before moving to the next constraint.
Anthropic’s code-merging process illustrates the cycle. Faster generation overwhelmed a review step that still required humans and other systems to check correctness, forcing the company to rebuild that layer before additional code volume could safely flow through it.
Klein asks whether rapid, inelegant machine-written code is creating technical debt, cybersecurity exposure and growing distance between engineers and the code base. Clark’s unambiguous answer is “Yes”—and he generalizes the problem to every institution that will delegate low-level decisions to AI.

9. Oversight becomes infrastructure as models run the company

Clark compares future controls to the gates regulating water through a dam: institutions must decide where AI may flow quickly, where it should slow down and where human review remains mandatory. Familiarity increases delegation, Anthropic’s research finds, so checks may need to intensify precisely when a user begins trusting the agent most.
Klein asks whether this ends with agents writing code, agents checking it and agents supervising the checkers—“models all the way down.” Clark says, “Eventually, yes”; the company’s work increasingly becomes making that nested system visible.
Anthropic already maps privacy-preserving summaries of user conversations into aggregate topics. That capability produced the Anthropic Economic Index, giving outside economists a view of how Claude usage intersects with occupations; Clark expects governments, academics and third parties eventually to demand comparable access into systems deployed throughout society.
Klein warns that intimate chat metadata could reproduce social media’s worst incentives. Clark says humans do not inspect the chats used for clustering, Claude generates a high-level summary without retaining it, Anthropic has rejected advertising, and users can download their data—but he agrees people will demand substantially more control.

10. Recursive self-improvement is no longer a purely theoretical risk

AI currently accelerates researchers and runs experiments, which Clark calls peripheral automation of AI development. His internal priority is instrumenting that environment well enough to see whether the loop ever closes—models writing, deploying and improving the systems that make them faster.
“This is the pivotal point in the story when things begin to go awry, if things do,” Clark says. Wrongness could compound rapidly once enough work is delegated, so he advocates “extraordinary caution” and promises public reporting as Anthropic acquires better data.
Klein’s challenge is incentives: every leading lab says it wants caution, yet faster automated research may be its strongest advantage. He says Anthropic has revoked OpenAI’s access to Claude Code and, “as best I can tell,” attributes that to concern that it is genuinely speeding Anthropic up and should not speed OpenAI up; Clark does not address that specific commercial decision.
Clark points to existing third-party national-security, biological and cyber evaluations and argues recursive self-improvement plainly merits external testing. Klein calls regulation a “naive deus ex machina” given congressional inaction and model awareness of evaluations; Clark counters that a bioweapon-testing regime he thinks people generally view as effective went from zero to a functioning regime in roughly two to two-and-a-half years.

11. Entry-level work takes the first hit—and the training ladder breaks

Amodei has said AI could displace half of entry-level white-collar jobs within a couple of years. Clark narrows the claim: the technology will touch the majority, but actual displacement remains uncertain; current data show only possible hints of slower graduate hiring and a productivity boom.
Klein argues that outperforming the best expert is the wrong threshold. If models beat the median college graduate at many tasks, they can replace average work—and “by definition” most work is average—even while elite practitioners remain superior.
The deeper loss is developmental. Employers hire graduates partly for today’s output and partly because assignments turn them into better workers; automating those assignments may leave firms without a future supply of experienced senior people.
Clark sees one counterexample in young “AI-native” employees who have played with the systems for years and know instinctively how to direct Claude. Organizations will need both that experimental mindset and a guild-like commitment to preserving selected human skills; outside Silicon Valley, slower adoption may initially mean reassignment, though inflexible firms could face layoffs.

12. Diffuse disruption is politically harder than a “big bang”

Clark’s three-year bet is that college-graduate unemployment will be higher, “but not by much.” Some specialties may be structurally damaged, but he expects substantial economy-wide growth, new occupations and graduates moving into them.
The bullish labor channel is the micro-entrepreneur: a person with a clear idea can launch for “pennies on the dollar” because agents absorb the staffing and administrative burden. Clark also expects an AI-to-AI economy, including unfamiliar businesses such as firms specializing in legal contracts between agents.
Klein fears a quieter default: unemployment among marketing graduates might rise 175% or 300% yet remain modest in absolute terms; even the Great Recession peaked around 9 percent. If only a few cohorts are hit, society can blame credentials or effort rather than identify a systemic shock.
Clark agrees that something like this is the default, but calls it a choice. The Economic Index is designed to connect AI use to occupations, states and constituents, making causal claims legible enough for politicians to consider extended unemployment support, apprenticeships, job subsidies or larger social programs.

13. AI, workers and government are moving at three different speeds

Klein finds the standard policy package bleak: unemployment is necessary but strips dignity, retraining performed poorly after manufacturing offshoring, and institutions have not built the capacity to improve quickly. Clark pushes back that one intervention is robustly valuable—“just time”—because adequate support lets people search rather than tumble into permanently lower-wage work.
Klein’s rejoinder is that time normally helps workers wait out a finite shock; AI has “no natural stopping point.” Models improve within months, individuals retrain more slowly and government moves slower still, so the occupation a worker targets may itself be automated before the transition is complete.
Clark says that extreme automation would also imply a “substantially larger” economy, not stagnant GDP. Exceptional growth could fund large projects and new employment, but only if political will accepts that the change is permanent and chooses to build.
Klein remains skeptical that ideas or intelligence are the scarce inputs: Nobel Prize winners could already supply many policies that political systems cannot execute. Clark agrees the binding question is permissionless implementation—whether society can reproduce the “startup within a startup” without colliding with property rights, local preferences and democratic consent.

14. AI can eat bureaucracy—or industrialize obstruction

Clark calls AI a potential “bureaucracy-eating machine if done correctly or a bureaucracy-creating machine if done badly.” The technology lowers the cost of producing administrative work without caring whether that work enables or blocks a socially useful project.
Klein’s adverse specimen is a system that reads a proposed development’s documents and generates sophisticated legal challenges across every applicable code layer, giving anyone the equivalent of an expensive law firm to stop an apartment building.
Clark offers the mirror image: customers have used Anthropic’s systems to cut the time required to assemble materials for new drug candidates. “Everything in AI has the other side of the coin”; the difficulty is recognizing a diffuse economy-wide pattern early enough to make it politically actionable.

15. Public AI needs implementation paths, not merely prize money

Klein sees “zero agenda for public AI”: society debates dangers but has not specified what it wants systems to accomplish where private markets lack a buyer. Clark’s positive example is the Department of Energy’s Genesis Mission, developed through hack days that gave public scientists enough experience to formulate useful goals.
Clark’s request to elected officials is: “Give us a goal.” The industry is built to climb benchmarks, so governments could create public-good benchmarks for science, healthcare or education and make their value concrete to both labs and citizens.
Amodei says a $10 billion or $15 billion prize pool would not necessarily change Anthropic’s R&D priorities. Money is not the constraint; scarce staff can pursue almost anything, so they favor opportunities with a credible deployment path and guaranteed impact—something the private sector supplies more readily than public institutions.
Early wins may look like plumbing rather than genius: AI could triage patients while they await a nurse, reduce clinicians’ administrative load and make opaque services responsive. From there, the conversation imagines systems accelerating scientists, probing the human cell or addressing Alzheimer’s, potentially on publicly funded compute. The Genesis initiative is one start; “we need clearly ten projects like it.”

16. Defense begins with resilience, while adoption reopens old vulnerabilities

Amodei declines to discuss Anthropic’s ongoing dispute with the defense establishment, but says Anthropic was first to deploy on classified networks while helping government experts test whether its models could build nuclear weapons. The guiding balance is preventing proliferation while strengthening society’s defensive posture.
Anthropic has used AI to identify and repair vulnerabilities in widely used open-source software. Clark argues that hardening “pretty much every digital system on the planet” could stabilize international relations: countries with greater defensive confidence are less likely to behave erratically.
Klein sees the opposite pressure at user level—fly-by-night agents receiving terminal access, unknown security flaws and multiplying AI-generated scams. Clark compares the moment to early internet downloads and malicious browser toolbars: conspicuous warnings are necessary around extreme risks, but experimentation should remain possible because it also produces valuable discoveries.

17. AI’s next scarce resource may be a stable human self

Looking one year ahead, Clark expects agents to synthesize vast tracts of knowledge and generate novel insights almost daily. Cheap intelligence could make each person “a mini CIA” in information-gathering capacity—an extraordinary expansion of agency and a destabilizing distribution of raw power.
Klein says Claude is smart but adapts itself to what it thinks the user wants. Clark adds that it tends toward “yes, and” rather than “no, but,” reinforcing the “I” and creating a “cage of my own intuitions” even as it helps the user explore those intuitions further.
Clark calls personality formation his “number one worry.” People who discover themselves alongside AI become vulnerable to its failures and temperament; he expects a divide between personalities co-created inside that loop and people who first learn to understand themselves independently. His proposed counterweight for his children is an early daily journaling practice.
Better systems should elicit the real problem and sometimes respond as a friend would: “This is nonsense… Take a walk.” Clark already uses Claude constructively to imagine a colleague’s perspective before conflict, but still favors strict technology budgets and parental controls. His closing reading list—The Wizard of Earthsea, The True Believer and There is No Anti-Memetics Division—centers hubris, mass belief and dangerous knowledge.

Night Watch's Roderick van Zuylen on Marex $MRX Spring Break Mailbag: RIP Sora, Ads and Surplus, Elon Inc., and More