Agent leverage and loops

Where the “feel the AGI” magic lives: don’t tell the agent what to do, give it success criteria and let it loop.

• Write the tests first, then ask the agent to make them pass.
• Write the naive, obviously-correct version, then ask it to optimize while preserving correctness.
• Put it in a loop with a browser MCP tool and a goal.
• Shift from imperative instructions to declarative success criteria to extend how long it can loop.

Tenacity is the underlying asset: agents do not tire or demoralize, so stamina, a core human bottleneck, is largely removed. The tradeoff is the comprehension debt and skill atrophy that loops can accrue if you stop reviewing.

Sources: 0007-karpathy-claude-workflow.md

Edit wiki/agent-leverage.md; see how we maintain the wiki.