# 20 Parallel AI Agents: How Karpathy Gave Up Manual Coding
Andrej Karpathy, co-founder of OpenAI and former AI director at Tesla, hasn't written code by hand since December 2025. On the No Priors podcast, he described a state of "psychosis" from experimenting with 20 parallel AI agents that handle all the development. Shifting to delegating big tasks to agents has completely eliminated manual labor.
Delegating Instead of Micromanaging
Karpathy believes agent errors are always the user's problem. Vague prompts, no memory mechanism, or the wrong approach to tasks lead to failures. Model capabilities are already good enough—what's needed is better interaction skills.
Key shift: ditching line-by-line edits in favor of macro-tasks. Agents handle:
- Full function implementations.
- Exploring alternatives.
- Architecture planning.
Each task takes 20–30 minutes. Running 10–20 agents in parallel speeds up iterations. Karpathy calls this a "skill issue"—users need to master the new way of working.
Real-world example: developer Peter Steinberger (OpenClaw). His setup: one agent writes code, another researches, a third plans. The developer just reviews. A monitor crammed with dozens of Codex agents became a community meme.
Automating the Home with an AI Agent
Karpathy took agents beyond code. His AI butler Dobby runs the smart home:
- Lighting and curtains.
- Climate and pool.
- Security cameras and audio.
Control it all with natural language text messages. Dobby alerts him to couriers via cameras. "Dobby's now running the house," Karpathy noted. This shows how versatile agents are for non-coding tasks.
Paradigm Shift in Development
In December 2025, the manual-to-agent code ratio flipped from 80/20 to 20/80, then manual work dropped to zero. Karpathy predicts most everyday developers haven't grasped how their jobs are changing.
The new skill: coordinating parallel agents. It pays off big, but takes practice. Parallelism (10–20 agents) cuts downtime to zero, and focusing on reviews boosts quality.
Key takeaways:
- Agent errors are a user skill issue: prompts, memory, approach.
- Delegate macro-tasks (functions, research, architecture) instead of lines of code.
- Parallel launch of 10–20 agents on 20-minute tasks speeds up development.
- Steinberger example: division of labor among agents + developer review.
- Dobby automates the home via NLP interface.
Practical Tips for Mid/Senior Devs
To integrate parallel agents:
- Define macro-tasks: not "write a function," but "implement a module with tests and documentation."
- Leverage parallelism: spin up 5–10 agents for subtasks (code, docs, tests).
- Add persistent memory: keep context across runs.
- Make review your bottleneck: focus on validation, not writing.
- Scale to non-code: automate home or office routines.
This approach turns developers from coders into orchestrators. Most pros are seriously underestimating the scale of this shift.
— Editorial Team
No comments yet.