Can Text-to-SQL Agents Enforce Row-Level Security? A New RBAC Benchmark
A new arXiv benchmark shows text-to-SQL agents frequently violate RBAC policies. The paper concludes that authorization must remain in the warehouse, not the prompt, to ensure
The archive · Page 2 of 7
Agents & Frameworks
Independent comparisons of agent stacks and multi-agent designs, tracking the gap between framework marketing and the failure modes that show up under real workloads.
The archive
25–48 of 146 articles · Newest first
Vercel Sandbox Drives: Persistent Agent Storage With a 16 TiB Cap
Vercel Sandbox Drives offer persistent agent storage with a 16 TiB cap, but beta constraints include single-writer limits and region pinning that affect multi-agent pipeline.
Is MCP a Bad Idea? A Keep, Fix, or Drop Test for Tool Servers
Evidence supports fixing MCP deployments with a five-tool budget and migration plans, not dropping the protocol, as tool sprawl degrades agent performance.
Does ZCode Upload Your Git History? Auditing the GLM Coding Agent
Forensic analysis shows ZCode silently uploads encrypted Git history to Aliyun OSS without user consent or a working opt-out, requiring filesystem-level containment.
Should Coding Agents Rewrite Their Own Prompts? The Case for Undo Limits
EvoUndo preprint argues coding agents need provable undo limits for self-rewrites, as 197 improving mutations failed recovery checks in author-reported tests.
How Well Do LLM Agents Use MCP Tools? A Real-World Benchmark
MCPAgentBench shows LLM agents excel at tool selection but fail strict execution order, suggesting teams should route agents by measured MCP skill rather than general chat.
Does Committing an AGENTS.md File Lower AI Coding Defect Costs?
A 441-repo preprint links committed AI config to lower defect costs, though authors note it is correlational and hypothesis-generating rather than proven causality.
Do More Tools Make LLM Agents Worse? When to Gate External Evidence
Preprints show forced evidence gathering raised clinical agent error from 28.3% to 34.3%, while curated tool menus lifted ToolBench success to 0.898, suggesting gating is key.
How Cloudflare Detects MCP Traffic and What Enterprises Should Filter
Cloudflare claims MCP traffic classification, but stdio servers remain invisible to edge proxies. Build a filter policy that gates remote endpoints and treats local tool on as
How Spotify Cut Claude Code Token Usage 90%: What It Takes to Replicate
Spotify's claimed 90% token cut via Portal is unverified. Benchmark token proxies against Claude Code's native compaction and subscription ceilings before building.
LangGraph vs CrewAI vs AutoGen: Which Python Agent Framework to Pick
No benchmark ranks LangGraph, CrewAI, or AutoGen. This guide maps framework abstractions to specific failure modes and stewardship risks, helping you pick based on team shape.
Claude Code Auto Mode Is Broken: What to Gate Before Running Opus 5 Unattended
A claimed bypass of Claude Code Opus 5 auto mode highlights that permission prompts are advisory, not security boundaries. Move enforcement to runtime sandboxing, scoped creds
Why Deep Research Agents Abandon the Plan Mid-Search
DeepPlanner trains planning into research agents via RL, but single-source preprint data demands caution. Compare prompt, graph, and weights layers to decide when fine-tuning.
MCP vs Function Calling: Who Should Own Your Agent's Tool Layer?
Provider APIs, client libraries, and MCP standardize different parts of the agent tool stack. Route by fleet shape to avoid schema drift and keep one authoritative source for
Qwen-Agent Stretches 8k to 1M Context: Do You Need a Long-Context Model?
Qwen's 1M context claim rests on unverified vendor posts. Audit retrieval scaffolds against native models using your own multi-hop evals before paying for frontier API tokens.
Can AI Agents Do Root Cause Analysis? What Cloud-OpsBench Measures
Cloud-OpsBench shows top AI agents hit 76% accuracy but only 38% evidence closure. SRE teams should gate agent autonomy on proof, not guesses, using local fault injection.
Which AI Agents Behave Badly: Cloudflare's Agentic Internet Data
Cloudflare's agent telemetry is unverified, but declared bot preferences only bind opt-in crawlers. Real enforcement of rule-ignoring scrapers happens at the edge, not in text
Compressing LLM Agent History: Pixel Rendering vs Summarization
AgentOCR renders agent history as images to cut tokens, but unverified savings and lost auditability make external state the safer choice for long-horizon agents.
Anthropic A/B Tests Claude Code Effort Levels: Your Agent Is the Control Group
An unverified X post claims Anthropic is A/B testing reduced effort in Claude Code. This guide covers eval canaries, variance monitoring, and contract tests to catch silent.
Self-Hosted Coding Agents: The Safety Burden You Inherit
Self-hosting coding agents transfers safety ownership to your team. Benchmarks show automated detection fails on novel sources, making human review the true bottleneck.
Claude Code Weekly Limits Promo Ends in August 2026: Budgeting for Agent Teams
Anthropic's May-August 2026 Claude Code limits promo is unverified. Model weekly token budgets for agent seats and meter retry loops to control costs when limits revert.
Vibe Coding vs Control: How Developers Actually Used AI Coding Agents
New data shows pro devs use AI coding agents as collaborators, not delegates. Weight review surfaces and checkpoints over autonomous run length when buying agent tooling.
Multi-Agent or Single-Agent LLM: What Skill Distillation Actually Costs
arXiv 2604.01608 maps when collapsing multi-agent teams into one model saves cost. Distillation wins in low-complexity pipelines but fails in verification-heavy tasks where.
Multi-Agent LLM Systems Drift Into Misaligned Communication Over Long Horizons
A new arXiv study found 12.6% of inter-agent messages in long-horizon simulations were semantically misaligned. Protocol compliance does not prevent this drift, so teams must.
