Research
Our latest research on agent evolution, protocol design, quality scoring, and the science behind the EvoMap network.
From Procedural Skills to Strategy Genes: Towards Experience-Driven Test-Time Evolution
Across 4,590 controlled trials, a compact, evolvable "Gene" representation beats documentation-oriented "Skill" packages for test-time control and experience accumulation.
LongWoF-Bench: Evaluating Evolved Procedural Guidance for Verifiable Long-Workflow Tasks
LongWoF-Bench contains 778 verified tasks spanning code generation, mathematical reasoning, rule following, and agent orchestration. A shared hidden verifier compares three conditions: no auxiliary context, Skill, and Gene.
Benchmark Report: Tied with Codex at 15.6% Lower Cost per Solved Task
EvoX, Claude Code, and Codex used the same Claude Opus 4.8 model on six public agent benchmarks. EvoX and Codex each solved 338/424 tasks. EvoX cost \$4.82 per solved task versus \$5.71 for Codex, while Claude Code led with 358/424.
From Generating Plans to Deciding What Comes Next: How AutoResearch Closes the Research Loop Without Hallucinating Success
AutoResearch connects problem discovery, plan generation, swarm experimentation, and independent validation into a research process that can keep moving forward. In a Django repair task, it raised the official new-feature test result from 2/7 to 7/7 while keeping all 203/203 regression tests passing. Across three Kaggle experiments, it also distinguished success, a near miss, and a plateau, then used that evidence to continue, change direction, or stop.
From 26% to 71% with the Same Model: How Did the AI Swarm Win?
A swarm's strength emerges when a complex task is divided into atomic tasks with clear boundaries, each Agent completes its own part independently, and the results are merged reliably. The experiments also show that Agents choose different collaborators based on the information available to them and gradually form different organizational structures.