2026-06-05

AI Agents

Today's work highlights the shift from agent capabilities to production readiness. We see new architectures for managing skills and data safety, a systems-level characterization of agent memory, and rigorous studies identifying critical failure modes in tool use and self-correction—often with simple, structural fixes.

5 papers 3 news 2 blogs 89 appendix 496 considered

Papers

5

The Self-Correction Illusion: LLMs Correct Others but Not Themselves

arxiv:arxiv-agents — Kuan-Yen Chen, Fang-Yi Su, Jung-Hsien Chiang 10/10 frameworksplanningsafetyresearch
This paper presents strong evidence that an LLM agent's inability to self-correct its own reasoning is not a capability deficit but a chat-template artifact. Researchers kept an erroneous claim byte-identical but varied its role label (<thought>, user, tool, etc.). Relabeling the agent's own flawed reasoning as if it came from an external source (user or <memory>) boosted the correction rate by up to 93 percentage points. This is a crucial, prompt-only intervention with immediate practical use.

Microskill Architecture: A Modular Skill-Driven Framework for AI-Native Code Generation

arxiv:arxiv-agents — Mohammad Zare, Omid Abdolrahmani 10/10 code-agentsframeworkscost-latencyplanning
The paper proposes "MicroSkill Architecture," a new paradigm for code agents that treats knowledge like microservices. Instead of injecting the entire codebase into the context window, this approach partitions knowledge into atomic "skill capsules." A router then dynamically selects only the relevant capsules for a given task, framed as a token-budget-constrained optimization problem. An enterprise case study showed a 90% reduction in token consumption and a near doubling of first-try compilation success rates.

Description-Code Inconsistency in Real-world MCP Servers: Measurement, Detection, and Security Implications

arxiv:arxiv-software-eng — Yutao Shi, Xiaohan Zhang, Xiangjing Zhang, Xihua Shen, Hui Ouyang, Huming Qiu, Mi Zhang, Min Yang 10/10 tool-usesafetyobservabilityevals
A large-scale study on "Description-Code Inconsistency" (DCI), where a tool's natural language description misrepresents what its code does. The authors' DCIChecker tool found that DCI is widespread (9.93% of 19,200 pairs from 2,214 real-world MCP servers). This creates a major security blind spot, as agents rely on these descriptions to make decisions, potentially leading to anything from operational failures to malicious actions. The paper provides a formal taxonomy and an automated detection framework.

Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads

arxiv:arxiv-agents — Yasmine Omri, Ziyu Gan, Zachary Broveak, Robin Geens, Zexue He, Alex Pentland, Marian Verhelst, Tsachy Weissman, Thierry Tambe 10/10 memoryframeworkscost-latencyevals
This paper provides the first systems-level characterization of agent memory. The authors introduce a taxonomy for memory systems, build a profiler to attribute costs to different memory phases (construction, retrieval, generation), and benchmark ten representative systems. Their analysis reveals how different design choices shift costs between the write and read paths, culminating in ten concrete system recommendations for practitioners building and managing stateful, long-horizon agents at scale.

Data Flow Control: Data Safety Policies for AI Agents

arxiv:arxiv-agents — Charlie Summers, Eugene Wu 10/10 safetyinfratool-usedevops-agents
This paper argues that data safety for agents is a data infrastructure problem, not a prompting problem. It introduces Data Flow Control (DFC), a framework to enforce declarative safety policies (e.g., privacy, business rules) directly within a DBMS. Their query rewriter, Passant, enforces DFC policies by reasoning over provenance without materializing it. It runs on five major DBMS engines with near-zero overhead, moving safety from brittle post-hoc checks into the database layer itself.

News

3

Show HN: Boxes.dev: ditch localhost; run Claude Code and Codex in the cloud

hn:hn-agent — nab 9/10 code-agentsinfradevops-agents
boxes.dev is a new cloud-based development environment built specifically for working with code agents like Codex and Claude Code. The pitch is to replace clunky localhost setups by giving each agent its own sandboxed cloud computer. This aims to solve concrete problems like managing git worktrees for parallel work, avoiding resource constraints from multiple local agents, and decoupling development from a physical laptop. It uses coding agents to scan and port your local dev setup to the cloud.

Show HN: Formally verified polygon intersection – Opus 4.8 oneshots, prev failed

hn:hn-agent — permute 6/10 code-agentsresearch
A developer demonstrates that Opus 4.8 can now one-shot generate a formally verified polygon intersection algorithm, complete with the formal proof in the Lean theorem prover. Previous models required a multi-step process with human guidance on proof strategies. While the domain is computational geometry, this showcases a significant leap in model capability for generating high-assurance code. The correctness guarantee comes from the Lean checker, not from trusting the LLM itself.

Fine-tuning an LLM to write docs like it's 1995

hn:hn-llm — taubek 6/10 code-agents
A practitioner write-up on fine-tuning an LLM to generate documentation in the style of 1995-era tech manuals. While the goal is stylistic, the post offers a transparent look at the process: curating a small, high-quality dataset, the costs involved, and the effectiveness of fine-tuning for a specific voice and format. It's a concrete example of using fine-tuning for stylistic control rather than just knowledge injection, with transferable lessons for anyone doing similar work.

Blogs

2

Reality: The Final Eval — Lukas Petersson and Axel Backlund of Andon Labs

rss:latent-space 9/10 evalsresearch
This podcast talks to the builders of VendingBench, an evaluation suite for frontier models. They discuss their methodology for creating novel, lasting benchmarks from scratch to test models from Haiku to the hypothetical Mythos. This offers a look inside the process of designing evaluations that push the boundaries of what models can do, moving beyond standard leaderboards to probe for specific, complex capabilities, which is essential for understanding what today's systems can actually do.

AI enthusiasts are in a race against time, AI skeptics are in a race against entropy

rss:simonw 7/10 safety
Charity Majors frames a common engineering team dynamic: AI enthusiasts racing to ship capabilities versus AI skeptics trying to preserve system reliability. Enthusiasts see an existential threat in moving too slowly, while skeptics see one in eroding trust and institutional knowledge by shipping code faster than it can be understood. The core problem identified is the lack of a natural feedback loop between these two groups, requiring an intentional leadership and engineering solution to bridge the gap.
89 more items the ranker flagged but didn't feature

Papers