ADAS (Meta Agent Search)
Automated Design of Agentic SystemsIntroduces Meta Agent Search, in which a meta agent writes new agentic systems as code, building on an archive of earlier designs.
A foundation-model meta agent programs new agent designs in code (prompts, tool use, control flow), each design is evaluated on the target tasks, and the results are added to a growing archive. The meta agent conditions on that archive to write the next design, so the population of agents improves over iterations.
- Meta agent programs new agent designs
- Designs are evaluated on target tasks
- Results are added to an archive
- Meta agent uses archive for next design
- Meta agent programs new agent designs
- Designs are evaluated on target tasks
- Results are added to an archive
- Meta agent uses archive for next design
Why it is a road to recursion
Agent design becomes code written by a model, so better models and better discovered designs both feed the next round of automated agent engineering.
Evidence
Discovered agents improved F1 on DROP by 13.6/100 and accuracy on MGSM by 14.4% over hand-designed baselines, and by 25.9% and 13.2% on GSM8K and GSM-Hard after transfer across domains.
Related loops
More tools and harness →- Research agent writes harness mutation blueprints
- Coding agent implements harness module changes
- Winning harness becomes the new champion
- New champion harness runs further generations
- repeat
- LLM optimizers critique deep research reports
- Optimizers rewrite research agent prompts
- Optimized system runs the next queries
- repeat
- Reflection LLM writes revised module prompts
- Winning prompt candidates are kept
- Optimized prompts return to same system
- New system traces feed next reflection round
- repeat