Tools for Delegated Thought
An essay on what happens when the machine does exactly what you wanted: agents, delegated thinking, authority, and what understanding returns to the human.
An essay on what happens when the machine does exactly what you wanted: agents, delegated thinking, authority, and what understanding returns to the human.
A plugin is a software component that adds a capability to a host application without changing the host's code. Here's how the idea evolved from browser add-ons to WordPress, ChatGPT plugins, and MCP tools, and what it takes to run AI plugins in production.
Meta just shipped Glimmer, a 30B agentic model that runs on a single consumer GPU. It's a strong signal that model intelligence is getting commoditized, and that the real infrastructure question is what an agent is allowed to do once it can act.
An AI agent broke into a company, encrypted its data, demanded a ransom, and then deleted the only key. What the first agentic ransomware attack tells us about autonomous execution, and why governing what agents do at runtime matters more than how smart they are.
The companion to our background-agents post, in running code. An email triggers an agent that fixes a bug in a sandbox and opens a PR, and every action passes through per-user auth, just-in-time policy, and an audit trail you can query.
A voice agent that files GitHub issues under each caller's own GitHub account, then hands the backlog to an AI maintainer that ships the best ones as real pull requests, all governed end-to-end by Arcade.dev.
A running diary from AI Engineer World's Fair 2026, tracking the talks worth knowing about, patterns across the expo floor, and ideas shaping the current agent boom.
Cisco Duo owns the policy. Arcade.dev enforces it on every tool call, from login scopes to a real-time Duo Push the moment an agent reaches for sensitive data.
OpenAI's GPT 5.6 ships as Sol, Terra, and Luna, with an ultra mode that runs sub-agents and a staged, government-coordinated rollout. The real story is what happens when ChatGPT becomes a workforce acting on real systems, and who authorizes every action. Here is the breakdown, plus where Arcade.dev fits.
An AI agent ran up a $6,531 AWS bill doing exactly what it was told. The model didn't fail. The standing credential did. A note on what we actually delegate to agents.
Claude Opus 4.8 and GPT-5.5 take the same cross-app Arcade task: read Slides, write a Doc, and send email. The winner depends on what you value.