Daedalus Paper Bootstraps Agent Memory From Self-Generated Tasks, +15.9 pp Success Across AppWorld and t2-bench

Illuin Technology's Daedalus pairs an explorer agent that generates practice tasks with a solver that attempts them, accepting a heuristic into memory only after the solver succeeds with it in context. Across AppWorld, t2-bench and AutomationBench, Daedalus lifts mean success rates by up to 15.9 percentage points and pass@5 by 2.2x over a no-memory baseline, without requiring training tasks or oracle verifiers. The authors show heuristics transfer to agents from other model families.