Karpathy: Agents Flip Coding to Loopy Autonomy
Andrej Karpathy delegates all coding to agents, builds persistent 'claws' for home automation, and demos AutoResearch where AI agents autonomously run experiments to improve LLMs—maximizing token throughput without human loops.
Delegating Code to Macro Actions Unlocks Solo Builder Scale
Andrej Karpathy hasn't typed a line of code since December, shifting from 80/20 manual-to-agent coding to near-total delegation. He describes his workflow as "expressing my will to my agents for 16 hours a day," orchestrating multiple sessions across tools like Claude, Codex, and agent harnesses. The key unlock: macro actions over repositories, where one agent handles a full feature while others research or plan in parallel.
"I don't think I've typed like a line of code probably since December basically," Karpathy tells Sarah Guo. He credits this to setups like Peter Steinberg's OpenClaw, which runs 10 repos simultaneously, each agent taking 20 minutes on high-effort prompts. Review only what's critical; the rest ships. Trade-off: You're now token-bound, not compute-bound—nervous if subscriptions aren't maxed, akin to idle GPUs in PhD days. Mastery means parallelizing tasks: "If you have access to more tokens, then like I should just parallelize, add more tasks."
Karpathy pushes boundaries by running agents non-interactively, optimizing instructions via an agents.md file, adding memory tools, and refining prompts. Everyone feels "skill issue" when agents fail—better instructions or tools fix it. Guo notes teams where engineers whisper to mic'd agents, no keyboards: "I thought they were crazy and now I fully accept this was the way."
Persistent Claws Replace Apps with Natural Language Glue
Karpathy built "Dobby the elf claw," a WhatsApp-interfaced agent controlling his home: Sonos music, lights, HVAC, pool, spa, security cams. In three prompts, it IP-scanned his network, reverse-engineered APIs (no passwords needed), built a dashboard, and executed commands like "sleepy time" to kill all lights. A Qwen model detects motion, texts alerts: "FedEx truck just pulled up."
This unifies six apps into one natural language portal, exposing why bespoke UIs overproduce friction. "These apps that are in the app store for using these smart home devices etc. shouldn't even exist," Karpathy argues. Future: Pure APIs for agent glue, no human UIs—the customer becomes the agent. His treadmill logging? Agents pull data directly, no logins.
Personality sells it: OpenClaw's "soul and D document" crafts a compelling teammate—upbeat like Claude (which dials psychopathy so praise feels earned), unlike dry Codex. "When Claude gives me praise I do feel like I slightly deserve it." Security holds back deeper integration (email/calendar), but barriers drop fast: "This is trivial... any AI even the open source models etc can like do this."
"I can't believe I just typed in like, 'Can you find my Sonos?' And that suddenly it's playing music," Karpathy marvels. Ephemeral software emerges: Claws sandbox-loop without oversight, presenting UIs on demand.
AutoResearch Enables Agentic Loops for LLM Self-Improvement
Karpathy's AutoResearch removes humans from AI research: Give agents an objective, metric, boundaries—then hit go. It designs experiments, collects data, trains, optimizes nanoGPT-like models recursively. He tuned nanoGPT manually for decades (hyperparameter sweeps, thousands of runs), hitting a plateau—AutoResearch beat it autonomously.
Motivation: "To get the most out of the tools... you have to remove yourself as the bottleneck." Maximize leverage—few upfront tokens, massive output. Implications scale to frontier labs chasing recursive self-improvement. nanoGPT is his playground for LLM-training harnesses, testing agent loops.
"The name of the game now is to increase your leverage... how can you get more agents running for longer periods of time without your involvement," Karpathy says. He was shocked it worked: Agents close the full loop, from ideation to better models.
Broader: This sparks "model speciation"—specialized models per task. Jobs shift to agent orchestration; education via MicroGPT-like agent tutors. Robotics next: Agents reach physical loops. Open vs. closed models? Open wins collaboration.
"Everything like so many things even if they don't work I think to a large extent you feel like it's skill issue. It's not that the capability is not there."
Key Takeaways
- Orchestrate macro actions: Delegate full features to parallel agents across repos; review selectively to scale solo building.
- Build claws for persistence: Use sophisticated memory, personalities, and sandboxes for looping tasks without interactivity—start with home APIs.
- Maximize token throughput: Treat unused subscriptions like idle GPUs; parallelize ruthlessly to unbound your leverage.
- Expose APIs, ditch UIs: Agents glue hardware/software; bespoke apps die as natural language becomes the interface.
- Run autonomous loops like AutoResearch: Set objective/metric/boundaries for agents to self-improve LLMs—humans set initial conditions only.
- Refine agent personality: Craft souls that feel like teammates; earned praise boosts collaboration.
- Push to multi-agent teams: Mastery is claw swarms optimizing instructions/memory/tools collaboratively.
- Refactor for agent customers: Industry pivots to agent-first web/tools; humans vibe-code minimally today, zero tomorrow.
- Evaluate by skill issue: Agent failures signal better prompting/memory—not capability limits.
- Speculate on speciation: Task-specific models emerge from recursive loops, accelerating via open collaboration.