Skill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI

Summary

• Karpathy says software engineering flipped in December from roughly 80/20 hand-coding versus agent delegation to 20/80—and has since moved much further toward agents. He says he has “basically” not typed a line of code since, while his operating constraint shifted from available FLOPs to how much “token throughput” he can direct across parallel sessions. The investable change is not merely better code generation: engineers now manipulate repositories through large delegated actions, making human orchestration the bottleneck.

• Persistent “claws” point toward an agent layer that absorbs the interfaces of conventional software. Karpathy’s Dobby agent discovered his Sonos system, reverse-engineered its APIs, and unified lights, HVAC, shades, pool, spa, and security behind WhatsApp; six separate apps effectively disappeared. His conclusion—“the customer is not the human anymore”—implies pressure on bespoke UX layers and growing value in hardware, accessible APIs, identity, memory, permissions, and secure agent infrastructure.

• AutoResearch found improvements beyond an experienced researcher’s manual tuning when the objective is cleanly measurable. After Karpathy manually tuned his small GPT-training playground, one overnight loop found interactions he had missed, including weight decay on value embeddings and better Adam betas. The scaling thesis is straightforward: explore thousands of ideas on smaller models, verify improvements cheaply, and extrapolate the winners—while contributing researchers become idea suppliers rather than experiment operators.

• The recursive endpoint is not merely models improving models, but agents optimizing the research organization encoded around them. Karpathy treats program.md as a primitive description of roles, procedures, and risk tolerance; competing versions could be benchmarked, with models learning which organizational instructions produce faster progress. “The LLM part is now taken for granted,” followed by agents, persistent claws, multi-agent systems, and eventually optimization over their instructions.

• Autonomy remains concentrated in verifiable domains, leaving a material gap between spectacular capability and general intelligence. CUDA optimization is ideal because speed and correctness are measurable, but softer intent, judgment, humor, and knowing when to clarify remain unreliable; Karpathy likens an agent to “an extremely brilliant PhD student who’s been a systems programmer for their entire life and a 10-year-old.” He expects more specialized intelligences, but says the science of changing weights without damaging capabilities is still immature.

• Karpathy is cautiously optimistic that cheaper software initially expands engineering demand through Jevons paradox rather than simply destroying jobs. Bits can be copied and manipulated vastly faster than atoms, creating a near-term “boiling soup” of digital refactoring; software that was previously too expensive becomes ephemeral and customized. He nevertheless refuses a confident long-term forecast and notes the uncomfortable endpoint: frontier researchers are “automating themselves away.”

• The discussion points toward closed frontier “oracles,” open models trailing by perhaps six to eight months, and a later wave connecting agents to the physical world. Karpathy prefers that balance to fully centralized intelligence, while imagining Internet-scale swarms whose untrusted compute could compete through expensive search and cheap verification. Robotics should lag because “atoms are a million times harder,” but sensors, laboratories, paid data collection, and marketplaces where agents purchase real-world observations could become the bridge to a potentially larger physical opportunity.

Deep dive

1. Coding shifted from keystrokes to repository-scale delegation

• Karpathy dates the discontinuity to December: his workflow flipped from roughly 80/20 writing code himself versus delegating to agents to 20/80, then moved further still. “I don’t think I’ve typed a line of code since December, basically,” he says, describing a change so abrupt that people outside software—including his parents—do not grasp how completely the default engineering workflow has changed.

• Sarah Guo recalls finding him determined to “express my will to my agents for 16 hours a day”—“manifest” being more accurate than code. His response is “perpetual AI psychosis”: capability has jumped, the frontier of possible individual output is unexplored, and every clever workflow demonstrated online creates anxiety that he is not yet using the tools at their limit.

• The concrete operating model comes from Peter, whom Karpathy describes tiling a monitor with Codex sessions across roughly 10 checked-out repositories. With high effort and good prompts, each task takes about 20 minutes; the operator assigns noninterfering functionality, research, planning, and implementation in parallel, then reviews the output according to how much the code matters.

• Work therefore moves in “macro actions”: not a line or function, but an entire feature that transforms a repository. Idle quota now feels like an idle GPU during Karpathy’s PhD—if Codex capacity runs out, switch to Claude—and the relevant resource becomes “what token throughput do you command?” Failure feels like “a skill issue,” which makes the workflow both empowering and addictive.

2. Persistent claws add identity, memory, and emotional fit

• Mastery means moving up the stack from one interactive coding session to teams of agents, then to persistent “claw-like” entities that loop inside their own sandboxes without constant supervision. Karpathy distinguishes this from ordinary context compaction: the persistent system may retain richer memory, continue acting while the user is absent, and expose one durable identity rather than a sequence of disposable chats.

• In discussing Claude and claw-like systems, Karpathy says there are “at least five” good ideas. He highlights a crafted personality and likely SOUL.md document, a sophisticated memory system, playful presentation, and a single WhatsApp portal into automation. The persona matters because users expect AI to be an entity that remembers, not the raw technical primitive of “more tokens come out.”

• Karpathy finds Claude’s personality unusually effective and Codex “a lot more dry.” Claude also calibrates praise well enough that he feels he must earn it: half-formed thoughts receive a neutral willingness to implement, while ideas he independently considers strong draw more enthusiasm. That controlled sycophancy makes the model feel like an excited teammate rather than an indiscriminate flatterer.

3. Dobby shows why agents may erase entire application layers

• During a January “claw psychosis,” Karpathy built Dobby, an agent for his home. Asked merely to find his Sonos, it scanned the local network, discovered an unprotected system, researched and reverse-engineered the endpoints, then played music in his study after roughly three prompts. “I can’t believe I just typed in, ‘Can you find my Sonos?’”

• Dobby subsequently connected lights, HVAC, shades, the pool, spa, and security. “Dobby, it’s sleepy time” switches off the house; change detection on an exterior camera forwards selected footage to Qwen, which sends a WhatsApp image and natural-language alert such as, “A FedEx truck just pulled up.”

• Six separate smart-home apps effectively collapsed into natural language. Karpathy’s stronger claim is that many such apps “shouldn’t even exist”: vendors should expose APIs, while agents supply the intelligence and cross-system glue that no individual interface can offer. Even treadmill exercise logging should be an agent-readable endpoint, not another login and web flow.

• Sarah asks whether normal users will really construct such systems. They note that today the process is involved and requires design decisions, but Karpathy predicts the barrier becomes trivial within “a year or two or three,” producing ephemeral software on each user’s behalf. He has not granted Dobby email or calendar access, however; security, privacy, and the technology’s rough edges still restrain adoption.

4. AutoResearch removes the experienced researcher from the loop

• Karpathy’s starting principle is leverage: “remove yourself as the bottleneck.” Instead of supplying the next prompt or inspecting each result, define an objective, metric, and boundaries once, then let agents run longer with minimal human tokens. AutoResearch is the worked implication of that idea—not simply assistance, but an arrangement in which the researcher no longer advances the loop.

• His small-model training projects are a playground for the deeper question of recursive self-improvement: can LLMs improve LLMs? Karpathy had already tuned the repository through two decades of experience, thousands of training runs, hyperparameter work, and what Sarah calls “earned confidence.” He expected little from an autonomous overnight search.

• The loop nevertheless found improvements he had missed, including weight decay on value embeddings and insufficiently tuned Adam betas. The important mechanism was interaction: once one setting changes, another setting’s best value can move. An objective validation criterion makes patient automated search useful for combinations a researcher might not manually test.

• Karpathy extrapolates toward frontier labs exploring cheaply on smaller models, where most experiments can be automated, then extrapolating successful changes. Researchers may still contribute ideas, as can automated scientists reading papers and repositories, but every idea enters one queue; workers test them, winning changes reach a feature branch, and humans occasionally merge into main.

5. Research organizations themselves become optimizable code

• program.md is Karpathy’s “crappy attempt” to specify how an autonomous researcher should operate: what sequence to follow and whether to inspect architecture, optimizer, or other idea classes. Different instruction files should produce different rates of progress, making the research process itself another measurable object of experimentation.

• His broader abstraction is that an entire research organization can be represented as markdown describing roles, communication, and decision rules. One coded organization might hold fewer useless stand-ups; another might take greater risks. Once both organization and output are code, the system can compare them and tune not only a model but the institution improving it.

• Karpathy proposes a contest in which people submit different program.md files for identical hardware. They discuss giving the resulting data to a model so it can write a stronger version: inspect which experiments worked, infer what instructions generated them, and recurse. The stack becomes “layers of an onion”—LLM, agent, claw, multiple claws, instructions, then optimization over instructions.

• The hard boundary is evaluation. Faster CUDA kernels are ideal because behavior can remain identical while speed supplies an objective reward; without a reliable evaluator, “you can’t AutoResearch it.” Sarah warns that autonomous loops may herd toward and overfit their stated metrics; Karpathy replies that the same machinery could devise broader suites of metrics.

6. Jagged intelligence limits how far autonomy can safely run

• The models are still “bursting at the seams”: moving too far ahead can make the whole system net unhelpful. Karpathy simultaneously feels he is speaking to “an extremely brilliant PhD student who’s been a systems programmer for their entire life and a 10-year-old.” That combination produces impressive long-horizon work alongside absurd mistakes and expensive loops that should have been rejected immediately.

• He ties the jaggedness to reinforcement learning. Labs can reward code that passes tests or answers with verifiable outcomes, but nuance, unstated intent, and knowing when to ask a clarifying question are softer. On on-rails tasks, the model operates at “the speed of light”; outside them, the response meanders.

• His favorite diagnostic is the joke that persists across model generations: “Why don’t scientists trust atoms? Because they make everything up.” Models can now “move mountains” on agentic tasks yet repeat the same weak joke from three or four years ago because humor was not comparably optimized.

• Sarah’s pushback—worth keeping—is that this undercuts the claim that code or mathematical gains automatically generalize into broader intelligence. Karpathy agrees that some transfer may occur, but “not a satisfying amount.” Capabilities across society are not arriving for free; opaque models contain persistent blind spots determined by what laboratories can train and reward.

7. Specialized intelligences may break the model monoculture

• Sarah asks whether a monolithic interface still makes sense when capability is so uneven. Karpathy expects more “speciation in the intelligences,” analogous to animals occupying niches with overdeveloped faculties. A smaller model could preserve a competent cognitive core while specializing in a domain such as mathematics in Lean, improving latency and throughput for users who do not need a universal oracle.

• Today’s labs instead merge improvements back into one model because they cannot know what any consumer will ask. Specialization may appear first in high-value business partnerships with bounded tasks, particularly where serving a massive general model for every request is economically or computationally wasteful.

• The obstacle is that manipulating the “brains” remains an underdeveloped science. Context windows are cheap and reliably customizable; touching weights through fine-tuning or continual learning can degrade other capabilities. Until laboratories can specialize models without destructive trade-offs—and cheaply enough to justify it—compute scarcity alone may not produce much visible speciation.

8. An Internet swarm could turn untrusted compute into research capacity

• Parallel AutoResearch becomes more interesting when workers are strangers on the Internet. A contributor might submit a code commit claimed to lower validation loss; finding it could require 10,000 failed experiments, while a trusted party needs only one run to verify the winner. Karpathy loosely compares this to a blockchain whose blocks are commits and whose “proof of work” is the expensive search.

• The architecture requires untrusted workers collaborating asynchronously with trusted verifiers, because automatically running arbitrary submitted code is “very sketchy and dodgy.” SETI@home and Folding@home illustrate the relevant asymmetry: locating a low-energy protein configuration is hard, but checking a claimed solution is comparatively cheap.

• A global agent swarm might even “run circles around frontier labs”: laboratories possess enormous trusted clusters, but Earth contains more untrusted compute. People could buy capacity for an AutoResearch track they care about—perhaps a particular cancer—instead of donating only money. Karpathy entertains a “flippening” from wealth to controlled FLOPs, then explicitly retreats: “I don’t actually think that’s true,” but finds it revealing that money cannot always procure compute.

9. Digital labor changes first, while cheaper software may expand demand

• Karpathy’s jobs project visualized public Bureau of Labor Statistics data, including decade-scale outlooks made in 2024, to reason through which occupations manipulate digital information. He avoids claiming that those jobs must rise or fall; demand elasticity and task composition matter. His firmer prediction is simply that they will change because AI is a third manipulator alongside humans and conventional computers.

• Digital agents are “ghosts or spirit entities” without physical embodiment, and bits can be copied vastly faster than matter can accelerate. That produces a near-term “boiling soup” of digital rewrites and perhaps 100-fold efficiency gains before robotics catches up—an “upgrade to the nervous system of the human superorganism.”

• His practical advice is modest: keep up with tools that are currently empowering, and remember that jobs are bundles of tasks, some of which can now move much faster. He understands dismissal and fear, but declines to pose as an economist; the long-term employment outcome is uncertain and varies substantially by profession.

• For software engineers, Karpathy is “cautiously optimistic” locally. Software was scarce because it was expensive; lowering cost may unlock demand through Jevons paradox, as ATMs reduced branch costs and were followed by more branches and tellers. Code becoming ephemeral means users need not accept fixed tools or subscriptions, creating enormous work to “rewire everything”—even as “a thousand-something” researchers at the labs actively automate themselves away.

10. Frontier labs and open models form an uneasy power balance

• Asked why he does not conduct AutoResearch with frontier-scale compute, Karpathy argues that ecosystem-level work can have substantial impact. Financial alignment inside a lab constrains what employees can freely say, while institutional pressure shapes public positions without anyone “twisting your arm.” When stakes become high, an employee may advise the room but is “not really in charge” of the entity.

• His concession is substantial: outsiders’ judgment will drift because frontier systems and future capabilities remain opaque. He would consider entering a lab for a period of serious work, then leaving again; moving back and forth might preserve both technical contact and independence. His answer to Noam’s question is therefore deliberately unresolved: excellent work and impact are possible on either side, and Noam’s most impactful work could well be outside OpenAI.

• On open source, the discussion recalls a lag that moved from no meaningful alternative, to roughly 18 months, to perhaps six or eight months. Karpathy expects open models to remain behind by some number of months. Linux demonstrates industrial demand for a shared platform—he uncertainly recalls it running perhaps 60% of computers—but model training carries far more capital expense. Open models may absorb ordinary and local workloads while closed “oracles” handle Nobel-scale research or projects such as moving Linux from C to Rust.

• He considers the current arrangement “by accident…an okay spot”: frontier capability remains expensive and closed, while slightly delayed intelligence becomes broadly accessible. Complete centralization carries systemic risk, yet the closed frontier itself has recently concentrated among fewer leaders. Karpathy wants more labs; Sarah wants “more people in the room,” arguing that ensembles outperform individual models and should inform humanity’s hardest decisions.

11. Agents will eventually buy observations and reshape education

• Karpathy’s self-driving experience makes him wary of robotics timelines: many startups disappeared, while progress demanded years, capital, and conviction. “Atoms are a million times harder,” so digital unhobbling comes first, then interfaces of sensors and actuators, and only later broad physical autonomy. The physical market might ultimately be larger, but its time arrives later.

• Once agents process the existing digital overhang—all the papers and uploaded information—they must “ask the universe questions.” Sensors can mean cameras, expensive materials-science equipment such as that used by Periodic, biological laboratories, or humans paid to generate training data—to “feed the Borg.” Actuators may likewise be people completing physical tasks before capable robots are ubiquitous.

• Sarah imagines information markets where an agent pays perhaps $10 for a timely photograph from Tehran to inform predictions, betting markets, or stocks. Citing Daemon as inspiration, Karpathy sees humans becoming both sensors and actuators for automated systems—not necessarily serving one another directly, but responding to the machine economy’s priced demands.

• MicroGPT offers the educational counterpart: an entire LLM training algorithm in about 200 Python lines, including roughly 50 for the network, 100 for autograd, and 10 for Adam. Karpathy no longer sees much value in recording a fixed walkthrough; agents can explain any function repeatedly and at the learner’s level. Documentation shifts from HTML for humans to Markdown and curriculum “skills” for agents.

• Yet an agent could not invent MicroGPT’s final minimal form, despite understanding it afterward. Karpathy’s long-running obsession supplied the scarce “few bits”; the model can route and teach them with infinite patience. His closing rule is stark: “The things that agents can’t do are your job now”—and everything they already understand should increasingly be delegated.