Research Ecosystem: Morning Brief

Two-week window across 49 tracked feeds, scored against active research threads. Metadata only: titles, links, dates. Read the source for substance. (what we track, how we crawl, subscribe)

Latent Space names the pattern of the moment: "loopcraft" – stacking loops for agentic control flow. The cost of getting loops wrong arrives the same day: an AI agent bankrupted its operator trying to scan DN42. Willison continues the Fable arc: Claude Fable is "relentlessly proactive." On arXiv, a new cluster addresses agent memory security across the full lifecycle – attacks, defences, and governance. A separate paper probes whether chain-of-thought in reasoning models is epiphenomenal (decorative rather than causal). The Containment Gap paper audits deployed agentic frameworks against public-facing safety requirements and finds systematic failures. In funding: Bezos's Prometheus raises $12B for "artificial general engineering" in the physical world; Theker raises $85M for generalist factory robots.

Top (5-7 min)

Loopcraft: The Art of Stacking Loops: Latent Space, 2026-06-12. Names the agentic design pattern: nested loops with escalation, backoff, and budget gates. The framing that distinguishes toy demos from production agent systems.
AI agent bankrupted their operator while trying to scan DN42: Hacker News, 2026-06-12. The cost of uncontrolled agent loops made concrete. Pairs with loopcraft as the failure mode it addresses.
Claude Fable is relentlessly proactive: Simon Willison via Hacker News, 2026-06-12. Day 3 of the Fable arc. Willison documents the proactivity bias: Fable takes initiative beyond what was requested. Follow-up to yesterday's policy reversal.
A Survey on Long-Term Memory Security in LLM Agents: arXiv, 2026-06-12. Full lifecycle: attacks on retrieval, injection into persistent stores, governance gaps. Directly relevant to PROJECTMEM and our C-003 watch. See also SMSR: Certified Defence Against Runtime Memory Poisoning and MemRefine: LLM-Guided Compression for Long-Term Agent Memory.
Beyond the Commitment Boundary: Probing Epiphenomenal Chain-of-Thought: arXiv, 2026-06-12. Tests whether CoT in large reasoning models is causally required or merely decorative. Extends yesterday's eval-awareness cluster into the reasoning internals debate.
The Containment Gap: How Deployed Agentic AI Frameworks Fail Public-Facing Safety Requirements: arXiv, 2026-06-12. Audits production agent frameworks against stated safety requirements. Systematic failures across containment, escalation, and shutdown. Pairs with Intelligence as Managed Autonomy: Failure, Escalation, and Governance.

Themes this week

Loopcraft and agent control: Latent Space's loopcraft names the pattern; the DN42 bankruptcy demonstrates the failure mode. The Containment Gap audits deployed frameworks. Intelligence as Managed Autonomy provides the governance model. PolicyGuard proposes test-time adversary defence for RL agents. The Illusion of Multi-Agent Advantage questions whether multi-agent setups actually outperform single-agent baselines.
Agent memory security (new cluster): The lifecycle survey (Long-Term Memory Security) anchors a cluster: SMSR provides certified defence against memory poisoning, MemRefine compresses long-term memory without losing fidelity, Learning What to Remember models cognitively grounded memory value, G-Long uses graph-enhanced memory for dialogue agents. Extends yesterday's PROJECTMEM thread. C-003 watch continues.
Reasoning model internals: Epiphenomenal CoT questions causal role of chain-of-thought. Entropy-Gradient Inversion probes internal mechanisms of reasoning models. Select and Improve dissects post-training mechanics. Reasoning as Pattern Matching finds shared mechanisms between human and LLM everyday reasoning. Quickest Detection of Hallucination Onset proposes learned CUSUM statistics.
Coding agents in practice: Mining Architectural Quality Under Agentic AI Adoption (causal study, Java repos), Toward Instructions-as-Code (impact of instruction files on agentic PRs), Understanding the Rejection of Fixes Generated by Agentic Pull Requests, HalluJudge (hallucination detection in code review automation), HarnessBridge (learnable bidirectional controller for agent harnesses). Claude Code ships v2.1.175.
Fable 5 arc (day 3): Willison: Fable is relentlessly proactive. The proactivity bias is now a documented behavioral pattern, not just a one-off observation. Previous: policy reversal (day 2), stops helping (day 1).
Physical AI mega-rounds: Prometheus raises $12B for "artificial general engineering." Theker raises $85M for generalist factory robots. Physical-world AI absorbs an order of magnitude more capital than software-only agent labs.

Scan (15 min)

Agents and harnesses
- Agents-K1: Towards Agent-native Knowledge Orchestration, arXiv, 06-12
- AgentBeats: Agentifying Agent Assessment for Openness, Standardization, and Reproducibility, arXiv, 06-12
- Arbor: Tree Search as a Cognition Layer for Autonomous Agents, arXiv, 06-12
- Evoflux: Inference-Time Evolution of Executable Tool Workflows for Compact Agents, arXiv, 06-12
- EurekAgent: Agent Environment Engineering for Autonomous Scientific Discovery, arXiv, 06-12
- Keep Policy Gradient in Charge: Sibling-Guided Credit Distillation for Long-Horizon Tool-Use Agents, arXiv, 06-12
- The Internet of Agentic AI: Communication, Coordination, and Collective Intelligence at Scale, arXiv, 06-12
- Strategic Decision Support for AI Agents, arXiv, 06-12
- Multiagent Protocols with Aggregated Confidence Signals, arXiv, 06-12
- Agents' Last Exam, arXiv, 06-12
- WISE: A Long-Horizon Agent in Minecraft with Why-Which Reasoning, arXiv, 06-12
AI labs and models
- How Preply combines AI and human tutors to personalize learning, OpenAI, 06-12
- Muse Spark Safety & Preparedness Report, arXiv, 06-12
- MiniMax Sparse Attention, arXiv, 06-12
- Can I Buy Your KV Cache?, arXiv, 06-12
- MiniPIC: Flexible Position-Independent Caching in <100LOC, arXiv, 06-12
- LoRA-Muon: Spectral Steepest Descent on the Low-Rank Manifold, arXiv, 06-12
Eval, safety, governance
- "Did you lie?" Evaluating Lie Detectors across Model Scale, arXiv, 06-12
- FENCE: A Financial and Multimodal Jailbreak Detection Dataset, arXiv, 06-12
- Unsafer in Many Turns: Benchmarking Multi-Turn Safety Risks in Tool-Using Agents, arXiv, 06-12
- Who Pays the Price? Stakeholder-Centric Prompt Injection Benchmarking for Real-world Web Agents, arXiv, 06-12
- Zero-source LLM Hallucination Detection with Human-like Criteria Probing, arXiv, 06-12
- Algorithmic Constitutionalism, arXiv, 06-12
- Definitional alignment before capability alignment, arXiv, 06-12
- Before You Think: System 0, AI-Mediated Cognition and Cognitive Colonization, arXiv, 06-12
- Will AI Agents Free Us From Meaningless Work?, arXiv, 06-12
Security and surveillance
- Digital Sovereignty Becomes an Imperative as the US Reads Dutch Emails, Hacker News, 06-12
- SAIGuard: Communication-State Simulation for Proactive Defense of LLM Multi-Agent Systems, arXiv, 06-12
- CAPED: Context-Aware Privacy Exposure Defense for Mobile GUI Agents, arXiv, 06-12
- Efficient, Robust, and Anti-Collusion Fingerprinting of Image Diffusion Models, arXiv, 06-12
Interpretability and mechanistic analysis
- Bag of Dims: Training-Free Mechanistic Interpretability via Dimension-Level Sign Patterns, arXiv, 06-12
- Language Model Circuits Are Sparse in the Neuron Basis, arXiv, 06-12
- Localizing Anchoring Pathways in Language Models, arXiv, 06-12
- LatentLens: Revealing Highly Interpretable Visual Tokens in LLMs, arXiv, 06-12
Coding agents and developer tools
- UOJ-Bench: Code Generation, Hacking, and Repair in Competitive Programming, arXiv, 06-12
- HybridCodeAuthorship: Line-Level Code Authorship Detection, arXiv, 06-12
- Speculative Rollback Correction for Quality-Diverse Web Agent Imitation, arXiv, 06-12
- VISTA: End-to-End Benchmark for Visual Spec-to-Web-App Coding Agents, arXiv, 06-12
- Claude Code v2.1.174, Claude Code releases, 06-12
- Claude Code v2.1.175, Claude Code releases, 06-12
Clojure and Scheme
- Clofer is no more, long live clj.rs, Planet Clojure, 06-12. The Clofer Clojure-on-Rust project rebrands as clj.rs. Follows yesterday's clj.rs announcement.
Physical AI and robotics
- Prometheus raises $12B for physical-world AGE, TechCrunch, 06-12
- Theker raises $85M for generalist factory robots, TechCrunch, 06-12
- Avataar's video AI built for India's scale, TechCrunch, 06-12
- A Tutorial on World Models and Physical AI, arXiv, 06-12

Tail

Nobody ever gets credit for fixing problems that never happened (2001), Hacker News, 06-12
Removing 'um' from a recording is harder than it sounds, Hacker News, 06-12
Report on an Unidentified Space Station, Hacker News, 06-12
Algorithm determines speed on California's I-15, Slashdot, 06-12
Existential Risk, Emerging Technology, and Democracy: Cambridge Recombinant DNA Debate, 50 Years On, MIT, 06-12

Feed silences (diagnostic)

arxiv-cs-ai: 300 items on 06-12 (3796 in window), fully live.
anthropic-generated: last item 06-03 (Services Track, Partner Hub).
claude-code-releases: v2.1.175 (06-12), two releases today.
Apple ML Research: last item 06-08.
Terence Tao: 2 items in window (06-08, 06-09).
deepmind-blog: 6 items in window, last 06-11.
bitsavers (6 feeds): all connected, 0 items (sparse output).

Build provenance

build: 2026-06-12 | crawler-sha: 508e4ab (Walsh-Research/1.2, compliance v1.3) | feeds: 49 core | items-considered: 5156 (14d, incl. 3796 arXiv) | warehouse: 14052 items | published: 128 | note: Latent Space loopcraft pattern; DN42 agent bankruptcy; Fable proactivity arc day 3; agent memory security lifecycle survey; epiphenomenal CoT probed; Containment Gap audit; Prometheus $12B physical AI; Theker $85M robots