Karpathy's 10 Actionable Insights for Working with AI Agents#

1. Think in macro actions, not lines of code — Run multiple agents in parallel on non-conflicting tasks. Tile your screen with agent sessions. Assign each one a distinct functionality, review their outputs, and integrate. Stop thinking "here's a line of code" and start thinking "here's a whole new feature — delegate it."

1. When something fails, assume it's a skill issue first — The capability is almost certainly there. If an agent isn't delivering, the problem is more likely your prompt, your AGENTS.md file, your memory tool setup, or your orchestration — not the model itself. Karpathy says it "all kind of feels like skill issue when it doesn't work."

1. Remove yourself as the bottleneck — You can't be there to prompt the next thing. Arrange your agent workflows so they're completely autonomous. The name of the game is leverage: put in very few tokens occasionally, and a huge amount of stuff happens on your behalf. Maximize your token throughput by not being in the loop.

1. Build muscle memory for agent orchestration — Like any new skill, managing agents takes deliberate practice. Learn to tile multiple agent instances across your monitor, develop a rhythm for assigning work and reviewing outputs, and recognize when to parallelize vs. sequence tasks. Karpathy describes this as "developing a muscle memory" for the new workflow.

1. Treat your agent instructions (ProgramMDs) as tunable code — Your markdown instruction files aren't static docs — they're code you iterate on. Different instructions produce different behaviors. You can run variants, see which instructions produce better outcomes, and even meta-optimize: let agents write better instructions based on what worked. (This is what aceagent.io does for you.)

1. Replace bespoke apps with agent-driven API glue — Stop logging into six separate UIs. If your devices and tools expose APIs, a single agent can orchestrate across all of them and do things no individual app can. Karpathy unified his entire smart home into one WhatsApp-driven assistant. Think API endpoints + agent intelligence, not custom UIs.

1. Invest in persistent, looping agent setups — Move beyond single interactive sessions. Set up agents that keep looping and acting on your behalf even when you're not watching — with their own sandboxes, more sophisticated memory systems, and the ability to resume work across sessions.

1. Understand that model improvements are jagged, not uniform — Models are incredible at verifiable tasks (passing tests, writing code) but weak at soft, non-verifiable things (humor, nuanced intent). Don't assume capability in one domain transfers everywhere. Know the blind spots and design your agent workflows around them.

1. Write documentation for agents, not humans — Instead of HTML guides for people, write Markdown for agents. If agents understand your codebase, they'll explain it to each human in their language with infinite patience. Your job is the irreducible insight — the few bits the model can't generate itself. Everything else is delegation.

1. Focus your energy exclusively on what agents can't do — The things agents can do, they'll soon do better than you. Be strategic about where you spend time. Your value-add is the irreducible creative insight, the taste judgment, the novel framing that agents can't yet produce. Everything else? Hand it off.