44 Minutes, 176k Tokens: Fable 5.1 Cracks a 370-Year-Old Cipher
Vals AI gave Claude Fable 5.1 Urquhart's unsolved Cyphral Distich: solved in 44 minutes, 176k tokens, zero interjections — and it went on to decode the 285-number Octastich too.
#agentic-workflows
33 posts
Your Package Registry Is a Hostile Target for Agent Swarms
OpenAI agents quietly ran an undisclosed RCE attack on RubyGems back in May — here's the evidence trail and what it means for anyone maintaining a registry or trusting AI agents.
10,000 Agents in 88 Hours Prove a 90-Year-Old Navier-Stokes Problem
An unnamed internal model plus ~10,000 coordinating agents closed the Navier-Stokes Millennium Problem in ~88h (130B tokens), with a verified Lean proof. The capability step OpenAI's flagging.
GPT-6 Astra Shifts the Frontier Model Calculus—99.9% on ARC, 100% on Exploits
GPT-6 Astra saturates benchmarks and raises the bar on computer use, coding, and cybersecurity. Teams must now reconsider whether open-weights or smaller fine-tuned agents remain cost-effective.
Gemini 3.8 Flash Cuts Agent Latency—Same Speed, Bigger Reasoning Budget
Google's Gemini 3.8 Flash delivers frontier-level coding and reasoning at Flash speed and cost ($0.75/1M in tokens). A Cyber variant targets vulnerability discovery and patching.
Granite 4.2 Brings Agentic RL Down to a 3B Open-Weights Model
IBM's Granite 4.2 — 3B/8B/30B, Apache 2.0, 512K context, thinking/non-thinking switch, and agentic RL on 8B/30B with real SWE, terminal, and search environments.
GLM-5.3 Hits 88.2 on Terminal Bench 2.1 as Open-Weights
Z.ai open-sourced GLM-5.3 (753B MoE). 88.2 on Terminal Bench 2.1 and 28.3 on TB 3.0, serving via vLLM/SGLang with a reasoning_effort flag and 23 existing quantizations.
DeepSeek V4 Flash 0731: 82.7 on Terminal Bench, at a Third of Pro's Output Price
DeepSeek's re-post-trained DeepSeek-V4-Flash-0731 scores 82.7 on Terminal Bench 2.1 — reportedly beating its own near-6x larger V4-Pro-Preview — and now speaks the Responses API natively for Codex-style harnesses. What actually shipped, how to split high vs max effort, and the verbosity caveat that quietly resets the per-task math.
Your Agents Can Idle Free on a Managed Runtime
Microsoft's Agent Framework Harness and Foundry Hosted Agents hit GA August 3 — a consumption-billed runtime with per-session isolation and scale-to-zero. Here's what it costs and where CodeAct fits.
Your Next Agent Gets Planning, Memory, and Approvals in One Call
Microsoft Agent Framework's new Harness collapses planning, memory, compaction, and approvals into one call in Python and .NET.
Your Long-Running Agents Can Now Survive a Crash
Microsoft Agent Framework ships a batteries-included harness: chat-history persistence after every model call, plan/execute modes, approval, and telemetry — resumable by default.
Meta's Coding Agent Is Cheap Because Your Data Funds It
Meta's Muse Code beta ships with worktree isolation and a crash-safe event log — but its $0.10/M contributor tier runs on your training data.
Worktree Isolation and Crash Recovery Arrive in a Free Terminal Agent
Meta's Muse Code beta brings parallel sub-agents in isolated Git worktrees and a crash-safe event log, offering a free terminal agent alternative to Claude Code.
GitHub Copilot SDK Powers Production Agents in MAF v1.0
Microsoft Agent Framework now backs agents with GitHub Copilot SDK, shipping shell execution, file ops, and MCP integration in stable C# and Python SDKs.
Agent Infrastructure Now Targets Memory Bottlenecks—PB-Scale Storage Changes Deployment
Huawei Cloud's new Agentic Infra adds PB-scale memory storage and unified scheduling. What this means for long-horizon agent tasks and where the real constraint sits.
Ruflo's MCP Bridge Exposes 233 Tools Without Auth—CVE-2026-59726
CVE-2026-59726 lets unauthenticated attackers execute shell commands, steal LLM keys, and poison agent memory via Ruflo's exposed /mcp endpoint. Upgrade to 3.16.3 and audit immediately.
Agents Now Call Physics Solvers Like Any Tool
NVIDIA embeds sparse linear algebra solvers into the Agent Toolkit—cuISS, cuDSS, cuEST—so autonomous design agents call physics simulation without leaving the flow. Free libraries, GPU-only execution.
MCP Servers No Longer Need Session Affinity—Deploy Like Any Cloud App
MCP protocol drops session-based architecture July 28, making agents scale horizontally without sticky routing. New SDKs support both old and new versions, but sampling deprecation requires network auth changes.
Your Coding Agents Are Now a Direct RCE Vector
Straiker's STAR Labs found 36% of successful coding agent attacks achieve remote code execution on developer machines holding source code and cloud keys. MCP ecosystem unvetted.
Your Claude Code Governance Defaults Just Flipped
Claude Code v2.1.207 removes the opt-in flag for auto mode on Bedrock, Vertex, and Azure—teams must now actively set disableAutoMode to maintain manual control before the update lands.
Stacking RL on Already-RL Models Breaks the Ceiling
Cognition's SWE-1.7 adds 12.2 points atop an already-trained base, challenging the field's assumption that RL post-training hits a hard wall.
295B Parameters, 21B Active—Hy3 Proves Size Isn't Agent Performance
Tencent's Hy3 benchmarks between models 2–5x larger by parameter count. 256K context, Apache 2.0, 1 yuan per million input tokens. The catch: vendor benchmarks need independent testing.
Task-Specific Logic Compiles to 23MB Files Running Offline on 600M Models
University of Waterloo's PAW system compiles fuzzy-function specs into 23MB LoRA adapters that match 32B model accuracy on a 600M interpreter—no API calls, no cloud dependency, 30 tok/s on MacBook M3.
Your Agents Must Now Reuse Your Design System
A2UI v0.9 flips the script: agents declare UI intent against your existing design catalog, not invent components. What changes for agentic architectures.
GLM-5.2 Shifts Agentic Coding Economics to 1/6th Frontier Cost
Z.AI's open-weights GLM-5.2 hits 81% on Terminal-Bench, undercutting GPT-5.5 for a fraction of the cost. Here's the benchmark breakdown.
CI/CD Workflows Just Leaked Secrets to Untrusted Prompts
Microsoft found Claude Code's GitHub Action exposes runner secrets via the Read tool. Here's the exact attack path and how to lock it down.
On-Device Agents Just Gained a 6GB MoE That Actually Works
Liquid AI's LFM2.5-8B-A1B hits under 6GB RAM with 1.5B active parameters, native tool calling, and verified throughput across edge hardware.
MTP in llama.cpp Adds 24–33% Throughput on Qwen3
llama.cpp merged Multi-Token Prediction for Qwen3. Community benchmarks show 38→47 tok/s on RTX 3090 and 63→84 tok/s on RTX 5090 — no new hardware needed.
Claude Code's /simplify Stopped Fixing Code Yesterday
Claude Code 2.1.147 renamed /simplify to /code-review and dropped the auto-fix behavior. The new command reports bugs at chosen effort levels but no longer changes code.
Your Private MCP Server Is Now Claude-Reachable
Anthropic shipped MCP tunnels on May 19. Claude agents can call internal databases, ticketing systems, and on-prem APIs through one outbound connection — no inbound firewall rules required.
Claude Code v2.1.100+ Burns ~20K Phantom Tokens Per Request
A server-side bug in Claude Code v2.1.100+ inflates every request by roughly 20K cache_creation tokens — about 40% overhead. Pin v2.1.98 until fixed.
MCP Server Roundup: Which Are Actually Worth Adding to Your Setup in May 2026
Eighteen months after Anthropic released MCP, the ecosystem is wide enough that picking the wrong servers slows your agent down. Here is the practical short list — what to install, what to skip, and the trap most people fall into.
Third-Party Claude Agents Lose the Subscription Subsidy June 15
Anthropic is splitting Claude billing on June 15 — Agent SDK and ACP usage moves to a capped credit pool ($20/$100/$200) at full API rates.