Charlotte and Chicago, two dinners this week. Leaders from some of the country's leading banks and beyond joined us. When we asked who had agents in production, far more hands went up than even just a few dinners ago. The panel gave way to attendees trading their own agent stories, and that's where the best conversation happened. More cities coming soon.

Your Outlook agent could always read an inbox. Now it can clear it. We rebuilt our Outlook toolkit and shipped agent-optimized toolkits for Snowflake, Calendly, and Airtable. Every tool is built for how agents work and scoped to exactly what each user can do. New agent tools breakdown here: arcade.dev/blog/agent-optimi…

81% of teams have agents in testing or live use, while only 14% have full security sign-off. That gap is dangerous. We wrote a playbook for the people stuck between those two numbers. The six things to enforce at the action layer, how to map governance onto the systems you already run, and a readiness checklist you can score yourself against today: arcade.dev/blog/ai-agent-gov…

AGNTCon and MCPCon Europe wrapped in Amsterdam last week. @shubcodes was on a panel about the new MCP release with Sam Morrow from GitHub and Shaun Smith from Hugging Face. Sam on the server side, Shaun on the client side, Shub on what the release means for the teams deploying against it. Thanks for stopping by, and to the @AgenticAIFdn for hosting!

TOMORROW: Your agent queries a database and a customer's bank account number comes back. It gets stripped out before the model ever reads it. We're showing how tomorrow in a workshop with @mastra_ai, along with two other enforcement points on every tool call. 9am PT. Register: linkedin.com/events/75075712…

Blessed that @techgirl1908 VP of @AgenticAIFdn, and absolute GOAT is gonna be hosting our panel, where Shaun (@huggingface) Shub (@TryArcade) and of course me @github will be discussing challenges and opportunities with MCP today at AGNTCon + MCPCon Europe 🚀

AIO went from niche to a line on every marketing plan in about a year, and most of us are still working out what it takes. The instinct is to reach for the SEO playbook, but it doesn't work. An LLM answering a question writes its own search queries first, then answers from whatever pages come back so you lose at steps you've never had to optimize for. @torresmateo open sourced the tool we built to measure it: arcade.dev/blog/aio-tool-how…

Arcade CEO @TheMostlyGreat reveals the 2 opposite failure modes killing enterprise agents: give them too little access and they’re useless, give them your access and they’re too dangerous to trust. "The biggest problem that everybody runs into when they try to connect an agent to something else is how do they trust that the agent can do it. We typically see two failure patterns." "The first one is they give the agent its own identity. If I create an agent that can read compensation data in Workday, and the intern has access to it, can the intern have access to the CEO’s compensation? The answer is no, so you end up reducing what it can do to a place where it’s not that valuable." "The other failure pattern is you put an agent on your laptop, and then the agent can do things as you. You can delete all your emails, delete all your files, and if you’re senior enough in an organization, you can drop a database table." "Where we come in is we help control and govern: can this agent, on behalf of this user, perform this action on this system?"

Guardrails stop a rogue agent most of the time. Probably. Hopefully. "Probably" doesn't cut it in an enterprise. You need to know your agent won't send that payment, delete that record, or fire off that email. That's what AI agent governance is for. We put together videos, articles, and a downloadable playbook to help: arcade.dev/aiagentgovernance… 🐶 (paid in treats, not dollars)

The Arcade team is in Amsterdam this week for AGNTCon + MCPCon Europe. @shubcodes is joining a panel on MCP challenges and opportunities with Sam Morrow of GitHub, Angie Jones of the @AgenticAIFdn , and Shaun Smith of Hugging Face. Catch the panel Friday at 16:10 CEST in the Emerald Room, and find the rest of the team at booth S3 both days. Schedule: events.linuxfoundation.org/a…

A loan officer asks an agent to approve a $95K loan and double-check its work. Four things go wrong and the model catches none of them. That's the running example in a live online workshop with @mastra on 9/22. Build a Mastra agent doing real work, with Arcade enforcing policy before the model sees a tool, before the tool runs, and before the result comes back. Register: luma.com/mastra-z2sd

I'm at @imaginationxyz today at Google Bay View for a panel called “What Runs Without You.” It's about what small teams are building once software that used to need a person at every step starts running on its own, and where a human still has to stay in the loop. Come find me if you're around.

An agent spent a few hundred Clay credits on one of our team's behalf last week, and the only record of what it did is the one the agent wrote itself. He only used Grok Bot because Clay's MCP server couldn't do the job he needed. When your MCP server covers six actions and your interface has two hundred, people will hand a browser agent their credentials instead. arcade.dev/blog/grok-bot-not…

Say your agent needs to send an email with an attachment to someone outside the company. Scopes answer whether it can send email at all, as that user. They can't answer whether that user should be sending this attachment to an external domain. That takes a policy check before the action runs, and a record afterward of what went out and which policies were in place. Full webinar on demand, next one 9/16: arcade.dev/webinars/2026-09-…

Execution Tool Logs are now live in Arcade. Every tool your agents run traced: which tool, which user it ran for, every retry, and the error if it failed. Arcade already decides what your agents are allowed to do. This is the record of what they actually did. @avoguru and Jose Barrueta wrote it up: arcade.dev/blog/execution-to…

Your agent isn't failing because it can't reason. It's failing because it has too many tools and no clarity on which to use. Arcade’s River Stevens and @render on why tool selection is a retrieval problem, plus a live demo of Arcade as a Claude plugin: piped.video/watch?v=kiUTMCtQ…

OpenAI shipped support for Skills Over MCP on August 26. The spec it implements is still not merged. @torresmateo on what that means. The extension is good engineering, but OpenAI imports skills as snapshots rather than live resources, so the headline feature doesn't work in the largest implementation. Breakdown and video: arcade.dev/blog/skills-over-…

The worst ROI on an agent: a large token bill for work that never got finished. Paying a painter who does a great job is not a problem. Paying for a bad job, or never letting them in the house because you can't be sure what they'll touch, is. The question isn't just what you're spending on tokens; it's how much your agents are getting done. @TheMostlyGreat on how AI agent governance gets you there: arcade.dev/blog/approve-once…

If you've run an agent on anything long, you know this one. You send it off on a two hour job, come back, and it stopped ten minutes in waiting for you to authorize. Arcade now clears all of it up front. The agent maps out what it needs and you approve it all in one pass before it starts, scoped to exactly what the job needs. More on the blog: arcade.dev/blog/pre-authoriz…