Gather — 27 July 2026

Reverse-Engineering the Coding Agent’s Own Architecture #

• Architecture Without Architects: How AI Coding Agents Shape Software Architecture (Konrad, Adam, Terrenzi, Ayvaz — arXiv 2604.04990, April 2026) — Names the phenomenon “vibe architecting”: coding agents pick frameworks, scaffold infrastructure, and wire integrations in seconds, decisions nobody reviews as architectural. Identifies five mechanisms of implicit architectural choice and six prompt-architecture coupling patterns ranging from contingent (structured-output validation, likely to weaken as models improve) to fundamental (tool-call orchestration, model-independent); a worked demonstration shows prompt wording alone produces structurally different systems for the identical task.
• Inside the Scaffold: A Source-Code Taxonomy of Coding Agent Architectures (arXiv 2604.03515, April 2026) — Source-level architectural taxonomy built from 13 open-source coding-agent scaffolds, scored across 12 dimensions in three layers (control architecture, tool/environment interface, resource management). Finds five loop primitives — ReAct, generate-test-repair, plan-execute, multi-attempt retry, tree search — function as composable building blocks, with 11 of 13 agents combining multiple primitives rather than using one control structure.
• Dive into Claude Code: The Design Space of Today’s and Future AI Agent Systems (Liu, Zhao, Shang, Shen — arXiv 2604.14228, April 2026, v2 July 2026) — Reverse-engineers Claude Code’s public TypeScript source and compares it against OpenClaw. Traces five underlying human values (decision authority, safety, reliable execution, capability amplification, contextual adaptability) through thirteen design principles into concrete implementation: a permission system with seven modes plus an ML-based classifier, and a five-layer context-compaction pipeline sitting around a simple model-call/tool-run while-loop.
• Don’t Blame the Large Language Model: How Agent Harness Evolution Shapes Coding Agent Quality (Ben Sghaier, Li, Adams, Hassan — Queen’s University, arXiv 2607.03691, July 2026) — First controlled longitudinal study isolating scaffolding’s contribution to coding-agent quality: fixes the underlying model and varies only the harness across 35 sequential scaffolding releases. Finding: practitioners who see quality regressions after a scaffolding update routinely blame the model rather than the harness that changed — direct empirical support for last cycle’s “harness, not model, is the architecture” framing.

Architecture as an Agent-Navigable Artifact #

• Formal Architecture Descriptors as Navigation Primitives for AI Coding Agents (Jin — arXiv 2604.13108, April 2026) — Controlled experiment (24 code-localization tasks, Claude Sonnet 4.6) shows structured architecture-descriptor documents (module boundaries, symbol signatures, constraints, data flows) cut navigation steps by 33-44% (Wilcoxon p=0.009). Format barely matters — S-expression, JSON, YAML, and Markdown all hit 95% accuracy — but an auto-generated descriptor reaches 100% accuracy vs. 80% blind. A field study across 7,012 real Claude Code sessions found formal declaration correlates with a 52% reduction in agent behavioral variance.
• Context Matters: Evaluating Context Strategies for Automated ADR Generation Using LLMs (arXiv 2604.03826, EASE 2026 Research Track) — Curates sequential ADRs from 750 open-source repositories to test five context strategies (no-context, all-history, First-K, Last-K, RAFG) for LLM-generated Architecture Decision Records. A small recency window (3-5 prior ADRs) beats full history; conclusion is that ADR-generation quality tracks context engineering more than model scale — a concrete instantiation of last cycle’s ARCHITECTURE.md-as-guardrail finding, now with a corpus behind it.

Debt Beyond the Code: Cognitive and Intent Debt #

• From Technical Debt to Cognitive and Intent Debt: Rethinking Software Health in the Age of AI (Storey — University of Victoria, arXiv 2603.22106 / ACM Queue, March 2026) — Proposes a triple-debt model: technical debt in code, cognitive debt in the erosion of a team’s shared mental model of a system, and intent debt in unrecorded rationale. Argues that as AI absorbs more code-writing, these two harder-to-see debts — not technical debt itself — become the dominant risk to a system’s long-term architectural coherence.
• AI Architecture Drift: How AI Agents Erode Your Codebase (techdebt.guru) — Practitioner taxonomy of three drift mechanisms specific to agent-generated code: pattern divergence (three modules, three different data-access approaches, each locally reasonable), layer violations (business logic in controllers, queries in view models), and drift velocity — AI can break architectural boundaries across dozens of files in a single session, versus the months/years manual drift takes.
• Building ARCHITECTURE.md Files That Prevent AI From Making Silent Architectural Decisions (jaikora, dev.to) — Names the enforcement gap directly: individual AI-generated PRs look professional and pass review individually, but the architectural decisions living “in the spaces between the rules” (composition vs. inheritance, new service extraction, renamed conventions) are invisible to both linters and human reviewers, and only cumulative drift reveals the problem. Confirms this cycle’s ARCHITECTURE.md approach is manual and effortful, not automatable — a limitation, not a solved problem.

Multi-Agent System Architecture #

• Moving From Monolithic To Microservices Architecture for Multi-Agent Systems (Goyal, Bhasin — arXiv 2505.07838, May 2025) — Applies the monolith-to-microservices architectural transition explicitly to multi-agent systems, covering Agent Communication Languages, MCP, and the A2A protocol as the connective tissue. Dated (May 2025, predates this journal) but not previously surfaced by the keyword set — flagged as a keyword-coverage gap below.
• A Two-Dimensional Framework for AI Agent Design Patterns: Cognitive Function and Execution Topology (arXiv 2605.13850) — Classifies 28 named agent design patterns along two independent axes (what the agent is cognitively doing vs. how execution is topologically structured — hub-spoke, mesh, hierarchical). More agent-general than coding-specific, but gives the orchestrator-worker/hierarchical/mesh vocabulary this journal already uses a formal backing taxonomy.

Cross-links #

• Claude-Specific Expertise The “Dive into Claude Code” paper (arXiv 2604.14228) reverse-engineers the actual Claude Code source — its 7-component structure, ML-based permission classifier, and 5-layer compaction pipeline are primary technical material for that topic too.
• Vibe Coding Approaches “Vibe architecting” (arXiv 2604.04990) extends vibe-coding vocabulary specifically to the architecture layer — prompt wording alone was shown to produce structurally different systems for an identical task.
• AI Code Quality A related multivocal review, “Faster Code, Deeper Debt? … Technical Debt and Its Early Signs in LLM-Assisted Software Development” (arXiv 2606.14796), surfaced during this cycle’s drift search — more line-level than architectural, but shares the debt-accumulation frame with the Storey cognitive/intent-debt paper tracked above.

Meta-observations #

• Emerging theme: A cluster of four arXiv papers from a single month (April 2026: 2604.03515, 2604.04990, 2604.13108, 2604.14228) each treat the coding-agent scaffold/harness itself — not the underlying model — as the primary unit of architectural analysis, extending last cycle’s “harness, not model, is the architecture” convergence from practitioner framing into a small but real academic sub-literature.
• Keyword suggestion: Add “vibe architecting” and “ADR generation LLM” to search.keywords — both are now named, citable concepts (arXiv 2604.04990 and 2604.03826 respectively) that the current keyword set would not reliably catch on a re-run.
• Gap: Quantified, agent-specific architecture-drift measurement is still rare. Most drift content (techdebt.guru, dev.to ARCHITECTURE.md piece) is qualitative practitioner description; the Formal Architecture Descriptors paper (2604.13108) is the only hard-numbers finding this cycle, and it measures navigation efficiency, not drift itself — the two are related but not the same claim.
• Source to watch: techdebt.guru produced the most specific practitioner content on agent-driven architecture drift this cycle (pattern divergence / layer violation / drift-velocity taxonomy) — not in current preferred or include_noisy lists; worth a staleness/quality check next cycle before deciding whether to add it.
• Method note: The monolith-to-microservices-for-MAS paper (arXiv 2505.07838, May 2025) predates this journal but wasn’t caught by any prior cycle’s keyword set — a reminder that this journal’s keyword-based search may be missing older-but-relevant foundational papers, not just failing to catch new ones.