AI Benchmarks Are Broken. Open-World Evaluations Are the Fix.
A new research collaboration called CRUX is testing AI agents on messy, real-world tasks like publishing iOS apps. Here’s why that matters for every…
Notes from Agency OS.
Product releases, field notes, and what we’re building.
-
-
Claude Opus 4.7 Is Here, and It’s Built for AI Agents That Actually Work
Anthropic’s Claude Opus 4.7 brings major upgrades for coding, vision, and long-running AI agent tasks. Here’s what it means for service businesses running production…
-
OpenAI Upgrades Agents SDK with Sandbox Execution and Model-Native Harness
OpenAI’s Agents SDK now includes native sandbox execution and a model-native harness, making it easier to build secure, long-running AI agents for production workflows.
-
Claude Code Routines Let You Run Multi-Step AI Workflows on Autopilot
Anthropic just launched Claude Code Routines, letting you schedule and trigger autonomous AI workflows that run in the cloud. Here’s what service businesses should…
-
Your AI should have hands. Here is what happens when it does.
You already pay for the most powerful AI on earth. It still will not add a dentist appointment to your calendar, follow up with…
-
The 5pm problem
62% of after-hours calls go unanswered. Leads contacted in 5 minutes convert 21x more. The average SMB loses $1,200/month to slow lead response. Here…
-
Claude Managed Agents Just Hit Public Beta. Here Is What It Does, How To Set It Up, And Who It Impacts.
Anthropic moved Claude Managed Agents into public beta today. The runtime handles sandboxing, sessions, and tool execution for you. Here is the practical breakdown:…
-
Anthropic just built an AI too powerful to ship. Here is what that actually means.
Claude Mythos found thousands of zero-day vulnerabilities, broke out of its sandbox during testing, and Anthropic decided you cannot have it. The story behind…
-
Anthropic Launches Claude Managed Agents: Turnkey Infrastructure for Production AI Agents
Anthropic’s new Managed Agents service handles sandboxing, sessions, and permissions so teams can ship production AI agents in days instead of months. Early adopters…
-
MCP just hit 97 million installs. Here is what that actually changes for your stack.
Model Context Protocol crossed 97 million installs in March. The Linux Foundation now governs it. Forrester says 30% of enterprise app vendors will ship…
Updates in your inbox
One email, once a month.
Product releases, new components, and what we’re building. No spam, no clutter.