HyEvo: Why Your AI Agents Are Still Too Slow (And Too Expensive)

Current agentic workflows have a dirty secret: they route everything through LLM inference — even the boring, predictable stuff. Need to parse a date? Call the LLM. Format output? Call the LLM. Run a regex? Call the LLM.

HyEvo fixes this.

The paper introduces hybrid agentic workflows that mix two types of nodes:

• LLM nodes for semantic reasoning (where the model actually earns its cost)
• Deterministic code nodes for rule-based execution (fast, cheap, exact)

The system then uses an evolutionary strategy to auto-generate the best hybrid topology for any task — iterating on both structure and node logic using execution feedback.

Results: up to 19× cheaper and 16× faster than leading open-source baselines across reasoning and coding benchmarks.

The implication is bigger than it looks. We have been treating LLMs as universal compute when they are really expensive probability engines. HyEvo is the first serious framework for surgically offloading predictable operations out of the inference path — a pattern that will define the next generation of production agents.

→ arXiv:2603.19639