ResearchConstraint Decay: LLM Agents Lose 30 Percentage Points When Backend Code Must Follow Architectural RulesarXiv·high signalResearchers quantify 'constraint decay' — as structural requirements accumulate (ORM patterns, API conventions, DB schemas), LLM agent code-gen performance drops ~30pp in assertion pass rates. Weaker configs approach zero. Flask handled well; Django and FastAPI trigger failures. Data-layer defects (wrong queries, ORM violations) are the leading root cause.SourceSource pagearXiv↳ Follow the threadShared entity / Stack layerTool-Making Agents Compile Repeated SOP Steps Into Versioned Tools Before DeploymentarXivShared entity / Stack layerTTHE: Test-Time Harness Evolution Lets Agents Rewrite Their Own Scaffolding Mid-RunarXivShared entity / Policy dependencyPrismata Confines Cross-Site Prompt Injection by Treating Web Agents Like an XSS ProblemarXivShared entity / Stack layerToken-Flow Firewall Audits Persistent Agents at the Natural-Language BoundaryarXivShared entity / Stack layerWebSwarm Uses Recursive Multi-Agent Orchestration for Deep-and-Wide Web SearcharXivStack layer / Threat patternTRACE Watermarks Agent Trajectories Against Adversaries Who Control the LogsarXivPolicy dependency / Stack layerMake context compaction a trained policy, not an inference-time heuristic — CompactionRL lifts SWE-bench up to +7 pointsarXiv 2607.05378Policy dependency / Stack layerStrict data-flow tracking is the only defense that zeroes out ADI — spend its utility cost only on code execution and moneyarXiv 2607.05120