Taxonomy of Self-Evolving Agents

Shilong Liu organizes self-evolving AI agents by what changes: the artifacts they produce, the harness that guides their work, or the underlying model. Artifacts include code, algorithms and scientific findings. The harness includes prompts, memory, tools, skills and execution loops. Model learning changes model weights.

These categories help compare approaches that use similar language but improve different parts of a system. They can also overlap: a system may refine an output while improving the tools or model used to produce it.

Verifiable tasks: any task where the environment gives an easy reward, such as benchmark scores, leaderboards or GPU kernel performance. In these cases an agent can repeatedly propose, test and refine candidates.

Examples discussed in the session include AlphaEvolve, which found improved matrix multiplication algorithms, and FARS, which produced more than 160 papers in over 400 hours at a cost of nearly $200,000. Tasks in the physical world remain much harder.

Agents can improve their harness by revising instructions, retaining useful lessons in memory and creating reusable tools or skills. These changes affect how the agent approaches future tasks while keeping the underlying model fixed.

The session discusses examples such as GEPA for prompt optimization, ACE for evolving playbooks, and Alita for tool creation. The common idea is to turn experience into something the agent can reuse, rather than repeating the same reasoning or implementation work for every task.

Adding more memory, instructions and tools can make an agent’s context harder to manage. Information useful for one domain may distract from another, and a larger tool collection can make tool selection more difficult.

Specialist agents offer one approach: each agent receives the tools and context relevant to its tasks, while a router assigns incoming requests. The session discusses this direction through Alita-G, which reported stronger performance at lower token cost. Routing remains a challenge, so specialization is an approach to evaluate rather than a guarantee that more agents will perform better.

Model learning can use signals other than human-provided correct answers. These include model-generated pseudo-labels, agreement between candidate answers, self-play and feedback from an environment. Depending on the method, these signals can become training targets or rewards.

The quality of the signal matters. A confident answer or repeated agreement is not necessarily correct. Learning without gold-standard answers therefore still requires a way to assess whether the resulting model improves on the tasks that matter.

These terms describe related but different ideas. Continual learning traditionally studies how a model acquires new capabilities while retaining earlier ones. A central challenge is catastrophic forgetting, where learning a new task reduces performance on previous tasks.

The session also discusses a line of test-time training research in which sequence models perform gradient-based updates during inference. Self-evolving agents is a broader framing here: improvement can happen in outputs, the agent harness or model weights. Comparing methods requires identifying what changes and which feedback drives the change.

Start with the task the system is meant to improve and the metrics that represent success. Faster code, more accurate answers and better robot behavior require different evaluation methods. Improvements to prompts, tools or model weights should be assessed against that intended outcome.

In the Q&A, Shilong Liu emphasizes choosing a main metric and focusing on it, since maintaining performance across all other domains while improving the target is nearly impossible. Better performance on one benchmark does not establish improvement everywhere. Evaluating systems that generate their own tasks or feedback remains an open challenge.

For robotics, strong benchmarks already exist, such as BEHAVIOR-1K. Many other real-world domains, including physics, math, biology and medicine, are under-explored and need domain-specific benchmarks.

Shilong Liu recommends testing models on tasks from your own field. He evaluates new models on a private subset covering robotics, visual understanding and computer use.

Not entirely. Shilong Liu expects unified models trained on language, multimodal and robotics data to become the core of robot control, with general models acting as policies.

Efficient, fast and robust low-level control modules will still be needed, so large models and specialized controllers are expected to work together.

The boundary can be subtle, and the classification is meant as a thinking aid. A practical rule: the final goal, such as a paper, a GPU kernel, a robot policy or a correct answer, is the artifact. The harness is the tool used to improve that result.

If an agent rewrites its own code, the harness can itself become an artifact, but there is always a final goal behind the process.