karpathy-patterns

Constraint-based design provides 10 core principles for autonomous agents, including single-file constraints, ratchet mechanisms, and Git as a state machine. These patterns ensure safety and effective context management derived from real autonomous system projects.

Context budget management restricts information flow to prevent overload in autonomous agents. It pairs with single-file constraints and the program.md interface to maintain efficient operational boundaries for LLM agent systems.

Git serves as a state machine for autonomous systems by tracking state transitions reliably. This design principle enables ratchet mechanisms and branchless DAGs to manage agent progress safely and efficiently.

The program.md interface defines a simple interaction layer for dumb platforms, ensuring autonomous agents operate within fixed constraints. This design principle allows collaborative infrastructure without complex platform dependencies.

Ratchet mechanisms secure irreversible progress while fixed time budgets prevent infinite loops in autonomous agents. These design principles ensure safe constraint-based design and predictable execution for AI architecture.

Constraint-based design limits agent freedom to enforce safety, which may reduce flexibility in highly dynamic tasks. These design principles prioritize controlled execution through branchless DAGs and single-file constraints over open-ended autonomous exploration.