Writing
Essays in the language they were thought in.
One story, ten animated drawings: agents can already pay for you, but your money has to get through three doors before you'll let them.
The payment step works, but the risk isn't under control at any of three layers: agents can't see consequences, fraud models can't price the risk, and users haven't built trust.
Three decisions behind how Harvey scaled agent development: package capabilities as tool bundles, ban custom orchestration, and gate every launch on evals.
An agent that spends money should run every purchase to the last step, show the user the final price, and have no tool that can place the order.
From Karpathy's 630 lines of Python to Harvey's legal agent: the engineering behind letting AI optimize itself overnight.
Once AI has generated a 400,000-word knowledge base for you, that's when the real problem starts.
Same model, different runtime, and its benchmark rank jumps from the top 30 to the top 5. The model didn't get smarter. The harness got better.