Behavior-Centric AI Design

This argues that the dominant agent-centric design - where a large language model drives control flow, decides tool use, and chains calls - creates fragile, monolithic systems that are expensive and hard to scale. When LLMs probabilistically decide next steps, complex tasks require large context, repeated token-heavy reasoning, and layers of validators or sub-agents to catch omissions, producing distributed monoliths with high operational, security, and cost burdens. That brittleness degrades user trust as systems scale: what works for a designer fails for diverse users, and industry signals suggest many agentic projects will not deliver expected ROI.

The proposed alternative is behavior-centric design: build a library of reusable behaviors with typed, validated inputs/outputs, versioned schemas, and structured outputs that are automatically indexed. Each behavior owns its data and credentials, and the platform enforces contracts at runtime so behaviors compose deterministically into workflows. Frontier models are invoked only for narrowly scoped tasks rather than governing control flow; workflows orchestrate behaviors, populate arguments from structured data, and pick or compute results algorithmically. This yields smaller LLM surface area, lower token costs, clearer security boundaries, easier observability, faster development, and more predictable, maintainable systems at organizational scale.