Experience → Action → Evolution
Building continuously learning, self-evolving agent systems across digital and physical worlds.
Team website · Our approach · Research agenda · Selected work
Experience should compound. Action should improve. Completing a task is not the end of learning. We study how an agent preserves what it discovers, puts that experience to work, and uses outcomes to improve the next interaction.
Our unit of evolution is the entire agent system—not only the underlying model. Memory, executable harnesses, world models, interaction structure, and feedback are complementary parts of one learning loop. Proposed changes must be evaluated before adoption.
| Stage | Research goal |
|---|---|
| Experience · 经验沉淀 | Turn interactions into retrievable memory, reusable knowledge, and evidence that can inform future decisions. |
| Action · 行动生成 | Put experience to work through memory, harnesses, and world models in digital environments and the physical world. |
| Evolution · 反馈进化 | Use outcomes to propose and evaluate changes, then carry validated improvements into the next interaction. |
The goal is not simply to complete more tasks, but to make experience useful beyond the task in which it was acquired.
| Direction | Focus | Explore |
|---|---|---|
| Experience & memory | Carry experience across interactions through reusable knowledge, provenance, lifecycle management, and selective forgetting. | Mi-Memory |
| Harness evolution | Turn knowledge into behavior through composable prompts, tools, and control flow that can be tested and improved. | HarnessX |
| World modeling & action learning | Learn environment dynamics from experience to inform prediction, simulation, and action learning. | WIA · World-model survey |
| Learning from feedback | Trace outcomes back to decisions, extract reusable skills, and evaluate proposed improvements. | CAR-bench TRACE |
| Collective intelligence | Study how roles, communication, and higher-order interaction structure shape learning and action in agent societies. | Topological Collapse |
| Embodied intelligence | Connect perception, reasoning, and physical action through executable programs that evolve from rollout feedback. | HarnessPAI |
| Project | What it explores | Resources |
|---|---|---|
| HarnessX | A harness foundry built from reusable processors and bundles, paired with models and evolved through training. | Code |
| Mi-Memory | A composable, evolvable, multi-source memory system for daily assistant agents. | Code |
| WIA | World modeling and action learning across digital environments. | Website · Code · Demos |
| HarnessPAI | An evolving executable harness for Physical AI. | Website · Code · Paper |
| CAR-bench TRACE | Trajectory-driven skill-bank evolution for in-car agents. | Code & report |
| Topological Collapse | How higher-order interaction structure constrains collective intelligence. | Code & reproducibility |
| Digital Agents Meet World Models | A survey and curated research map for world models in digital-agent environments. | Survey & collection |
Explore our open-source projects, contribute to a relevant repository, or open a project issue to discuss research and engineering ideas.
The official team website is darwin-agent.github.io, maintained under the Darwin-Agent organization.