Back to AI Insights

We Are Still Writing Assembly for AI Agents

A thesis on why agents need language-level safety: typed limits, tool contracts, human approval gates, and compile-time constraints.

Every leap in computing created a higher-level abstraction for humans. LLM agents are another leap, but today we still hand-wire loops, tools, context, memory, safety, and budgets one fragile line at a time.

Frameworks help, but they do not make dangerous designs impossible. The missing layer is closer to a DSL or static analyzer: a system that refuses to run an agent loop without limits, a destructive tool without approval, or unvalidated structured output.

The first useful version does not need to be a full language. It can begin as TypeScript for agents: typed configuration, static checks, and production invariants encoded outside prompts.