Survey publication slice / Evidence-backed article

Slice 04 — Why Self-Evolving Agents Break After the Demo

Slice 04 — Why Self-Evolving Agents Break After the Demo is an indexable Self Evolve survey publication slice that turns papers, projects, benchmarks, and public evidence into an article asset with explicit review boundaries.

Slice 04 — Why Self-Evolving Agents Break After the Demo

One-Sentence Thesis

The gap between a convincing demo and a production self-evolving agent is mostly about observability, state, cost, rollback, and governance, not about one more clever prompt.

Reader Value

This slice converts the survey’s user pain points into a production checklist for teams considering agent self-improvement.

Three-Sentence Summary

Many agent systems look impressive when the environment is small, the objective is hand-picked, and the failure is manually filtered. Production introduces long horizons, tool side effects, changing context, cost constraints, security boundaries, and humans who need to debug what happened. A valuable survey slice should therefore show what breaks, how to instrument it, and what evidence is needed before trusting self-modification.

Evidence Chain

ClaimEvidence
The survey includes an industry chapter and a pain-points chapter.survey/latex/chapters/ch6-industry.tex, survey/latex/chapters/ch7-painpoints.tex
The English pain-points chapter maps reliability, debugging, production deployment, cost, latency, monitoring, and governance.paper-drafts/ch7-painpoints.tex
Social/blog mechanism analysis extracts real pain points from HN, Reddit, Chinese communities, and blogs.work/wiki/sources/raw-social-blogs-mechanism-analysis.md

Production Checklist

AreaQuestion
ObservabilityCan we replay the trace that caused a self-update?
StateWhere are memory, skills, prompts, and code changes stored?
EvaluationWhat prevents the agent from optimizing the wrong metric?
RollbackCan a bad self-change be reverted automatically?
CostDoes improvement pay for its own search/evaluation budget?
SecurityCan generated tools or code cross unsafe boundaries?
GovernanceWho approves persistent changes and under what policy?

Public Report Shape

  1. Open with “the demo is not the product; the loop is the product.”
  2. Give a production failure taxonomy.
  3. Map each failure to required instrumentation.
  4. Provide a minimal deployment readiness checklist.
  5. Link to project model cards that do or do not address these requirements.

HN Submission Candidate

Title: Why self-evolving agents break after the demo

This is likely useful for engineering communities and practical AI builders.

Counterexamples And Limits

Some research systems are intentionally not production systems. This slice should avoid judging pure research by enterprise deployment standards unless the project claims practical use.