이미지: X — 모델·오픈소스 화면 갈무리
Summary
- Tencent Hunyuan introduced a survey paper on the reliability of AI agent self-evolution via X
- It classifies the depth of change into five levels, L0 through L4, from output-level changes to changes in judgment criteria
- The paper argues that for self-evolution to be trustworthy, the evidence must hold up even outside the boundary of the given update
- 발표 채널
- X(트위터) Tencent Hunyuan 공식 계정
- 발표 시각
- 2026-08-12T07:42:34+00:00
- 논문 제목
- Diving into Reliable Self-Evolving Agents: A Survey
- 핵심 프레임워크
- 변화의 깊이를 L0~L4 5단계로 구분
- L0~L2
- 출력·모델·스캐폴드(프롬프트·스킬·메모리·워크플로) 수준의 변화
- L3~L4
- 향후 업데이트를 제안·관리하는 방식과 그 판단 기준 자체의 변화
- 핵심 주장
- 신뢰할 수 있는 자기진화는 관련 업데이트 경계 밖에 남는 증거가 필요하다
An agent says it fixed itself — but how can we trust that
If an AI agent finishes a task and then rewrites its own prompts, reorganizes its memory, or even changes its own evaluation criteria, what basis does a person have for judging that it has actually "gotten better"? On August 12, Tencent Hunyuan introduced via its X account a survey paper attempting to answer this question: "Diving into Reliable Self-Evolving Agents: A Survey." The work systematically organizes the ways agents modify themselves and maps out what evidence is needed to trust each update.
Breaking the depth of change into five levels
The paper's core framework classifies changes into five levels based on "how deeply the change penetrates." L0 covers output- and behavior-level changes that affect only the current task, while L1 covers changes to the trainable model or policy itself. L2 involves changes to the scaffold — prompts, skills, memory, workflows, and the harness (the execution framework wrapping the agent). L3 addresses who proposes and manages future updates and how, and L4, the deepest level, involves changing the very criteria used to judge behavior and future updates.
| Level | Name | Scope of Impact |
|---|---|---|
| L0 | Output-Level | Limited to the current task |
| L1 | Model-Level | Trainable model/policy |
| L2 | Scaffold-Level | Prompts, skills, memory, workflows |
| L3 | Improver-Level | How future updates are proposed and managed |
| L4 | Criterion-Level | Criteria for judging behavior/updates |
The paper's argument is clear. Changes that stay within L0 and L1 — confined to the immediate task — are relatively easy to verify, but from L2 upward, the change carries over into the next task and the next update. In an attached image, Tencent Hunyuan states plainly that "evidence must persist beyond the boundary of the update." Underlying this is a concern that if an agent goes so far as to change the very criteria by which it judges itself, it risks falling into circular reasoning — using its own new judgment criteria to claim its own change was correct.
Why this question is coming up now
Self-evolving agents have recently moved rapidly beyond the experimental stage across the industry. On August 8, Kimi unveiled "Agent Swarm," a system that runs up to 100 sub-agents in parallel and handles more than 1,500 tool calls. On August 10, Sakana AI expanded its RSI Lab and presented "physical AI" — systems that improve themselves in the physical world — as the next research frontier. The directions differ, but the underlying premise is the same: building systems in which agents improve their own performance without human intervention.
The problem is that the means to prove such improvement haven't kept pace. When a model rewrites its own prompts or restructures its memory, a rising benchmark score alone doesn't reveal whether that change holds up in other situations. Tencent Hunyuan's survey reads as an attempt to fill exactly this gap in evaluation methodology — namely, what evidence can be used to trust the results of self-evolution.
So what changes
The paper itself is not a new model or product. But for those building self-evolving agents, it offers a reference baseline. Rather than simply claiming "performance improved," the framework calls for first identifying which level, L0 through L4, the change occurred at, and then demanding evidence appropriate to that level. It could serve as a checklist for filtering out illusions — such as dressing up a scaffold-level change as a model-level improvement, or changing the judgment criteria itself and then claiming things got better. For companies looking to deploy agents in real services, it adds one more question to ask before simply accepting a claim that a system has "gotten smarter."
