Evolving Excellence: Automated Optimization of LLM-based Agents
Figure 1: The agent as genome. Prompt, tool, and parameter genes recombine under semantically-aware operators (a); selection on measured accuracy and token cost drives the population toward cheaper, better agents (b).∎
An LLM agent is a bundle of decisions — system prompts, tool wiring, model parameters — usually tuned by hand and rarely revisited. This paper treats that bundle as a genome. Artemis applies semantically-aware genetic operators: crossover and mutation that respect the meaning of prompt fragments and configuration, not just their text.
Populations of agent variants compete on real workloads spanning competitive programming and mathematical reasoning. Selection favors accuracy and cost together, and the winning configurations transfer without any architectural change to the agent.
Key results
- ≈37% token-cost reduction with answers preserved
- Accuracy gains across programming and mathematical-reasoning workloads
- No architecture change — only prompts, tools, and parameters evolve