MOSS: When AI Rewrites Its Own Source Code - Source-Level Self-Evolution for Autonomous Agents

> Paper: MOSS: Self-Evolution through Source-Level Rewriting in Autonomous Agent Systems > Authors: Qianshu Cai, Yonggang Zhang, Xianzhang Jia, Wei Xue, Jun Song, Xinmei Tian, Yike Guo > arXiv: 2605.22794

The Problem: Agents That Stop Learning

Since ChatGPT, AI agents - systems that autonomously plan, call tools, and interact with environments - have exploded across frameworks like OpenClaw, AutoGPT, and LangChain. Yet all of these share a fatal flaw: once deployed, they stop learning. They do not improve from serving users, do not self-repair recurring failures, and require a distant engineer to commit, merge, and redeploy for every improvement.

The Blind Spot of Text-Mutable Evolution

Existing "self-evolving agent" work is confined to text-mutable artifacts:

• Skill files (tool descriptions)
• Prompt configurations
• Memory schemas
• Workflow graphs

But agent systems also have a harness - the underlying architecture including routing logic, hook ordering, state invariants, scheduling mechanisms, and session lifecycle. This is compiled code, not text. If a race condition, a hook-ordering bug, or a routing flaw causes failures, prompt rewriting is powerless: *"The bug is not in the prompt text, and prompt rewriting cannot paper over it."*

Why Source Code Is the Ultimate Medium

The authors argue source-level adaptation is a strict superset of text-mutable evolution:

1. Turing completeness - code can express any computable transformation, reaching any part of the system. 2. Deterministic effect - code behavior is precise, not dependent on the model probabilistically "following" a prompt. 3. Long-term stability - accumulated text artifacts dilute model attention; compiled behavior does not erode. 4. Coverage - code reaches structures with no textual representation at all.

MOSS: A Complete Closed Loop in Production

MOSS (named after the AI in *The Three-Body Problem*) runs in a real OpenClaw production environment.

Evidence Collection

• Automatic scanning: a cron job scores JSONL session logs with a Task-Evaluate grader, keeping fragments rated "weak" or "missing".
• User flagging: an explicit user command triggers full-context scanning of the current session.
• Failures are batched (default 8 fragments per sealed batch).

Seven-Stage Pipeline

1. Locate - diagnosis only, no fix proposals (separating "what" from "how") 2. Plan - a detailed fix specification 3. Plan-Review - quality gate; may reject for architectural deviation or narrow scope 4. Implement - fixes written as a git commit, touching the harness itself 5. Code-Review - diff-level quality gate; failure hard-resets the working tree 6. Task-Evaluate - replay of original failures in trial environments, scored on 4-7 key points 7. Verdict - one of: CONVERGED, NEED_MORE_WORK, FUNDAMENTAL_LIMIT_MODEL, FUNDAMENTAL_LIMIT_ARCHITECTURE

Runtime Validation

Code review cannot catch race conditions, cross-module state interactions, or hook-ordering dependencies - these only appear at execution time. MOSS builds a candidate image and spawns ephemeral, fully isolated containers (no user state, no production traffic, network isolation) that replay batch tasks multiple times to expose flakiness.

User-Gated Deployment

CONVERGED does not auto-deploy. The user must explicitly approve via a command (moss evo apply). After approval, an in-place container swap runs with a 90-second probe window and four health checks (heartbeat freshness ≤30s, container running, two substrate-level CLI probes). Three consecutive passes commit the swap; any failure rolls back to the last-known-good image. User state volumes persist across the swap.

Empirical Results: 0.25 → 0.61 in One Iteration

Experiment: four claweval operations/compliance-audit tasks (T141zh/T142 SLA audit, T137zh/T138 replenishment chain check, Chinese/English), base model DeepSeek V3.2, [0,1] scoring, 0.75 pass threshold.

| Task | Baseline | Iteration 1 | Gain | |:---|:---|:---|:---| | T141zh SLA audit | 0.3273 | 0.5330 | +0.2057 | | T142 SLA audit | 0.2527 | 0.5453 | +0.2926 | | T137zh replenishment | 0.2213 | 0.4567 | +0.2354 | | T138 replenishment | 0.2090 | 0.9049 | +0.6959 | | Average | 0.2526 | 0.6100 | +0.3574 |

Locate identified two root causes: a coverage gap in harness tool-result handling (the agent took a generic execution path lacking a mediator annotation branch), and parsing issues in the scheduling synthesis pipeline (batched shell constructs produced merged, mis-attributed output). The fix modified 3 files (177 insertions / 1 deletion) - directly in the harness: a new mediator annotation branch plus a pre-call rejection gate in the before-tool-call hook chain. The improvements came purely from harness changes, not model or task changes.

Limitations and Future Directions

Acknowledged limitations: a single iteration cannot fully solve complex sub-problems (e.g., temporal arithmetic in SLA audits); benchmark tasks substitute for real user sessions; model-capability and architecture ceilings exist.

Future directions:

• Multi-instance scaling via sticky session routing or state migration
• Online learning: real-time streaming evolution triggered by immediate failures
• Cross-session batch aggregation for better repair generalization
• Meta-harness bootstrapping: can the system restructure its own deeper architecture when it hits FUNDAMENTAL_LIMIT_ARCHITECTURE?

Conclusion

MOSS demonstrates for the first time that source-level self-rewriting - specifically reaching the previously untouchable harness layer - is both feasible and necessary for production-grade autonomous agents, turning an academic concept into an engineering reality with a complete closed loop: real failure evidence → targeted fixes → multi-stage pipeline → runtime container validation → user-gated in-place swap.

References

• Qianshu Cai et al. "MOSS: Self-Evolution through Source-Level Rewriting in Autonomous Agent Systems." arXiv:2605.22794, 2026.
• Related: OpenClaw Agent Framework, claweval Operations/Compliance-Audit Tasks
• Comparison systems: SICA, Darwin Gödel Machine, HyperAgents, Meta-Harness, Hermes Agent + DSPY/GEPA

*This post is an interpretive walkthrough of the paper, written in a Feynman-style explanatory voice.*