Chris Parsons

I've split my AI work into little bots with an avatar and personality, each doing a few things. I talk to them one at a time, a bit like I talk to different employees. That doesn't make them people, and I'm not suggesting we pretend it does. The conversations just feel familiar. And, honestly, they're more fun this way. Not sure if that’s just because it matches an out of date mental model though. I think we'll anthropomorphise AI agents whether we mean to or not. What do you think the right way to engage with AI is?

I built a Pi extension to block certain irreversible actions - I'm a big believer in hard rules not instructions. However, when I drew six of my own workflows in AWML. One showed posts being scheduled on an approval that wasn't recorded (oops). AWML (Agentic Workflow Modelling Language) shows me where the approval should happen, and where content crosses a boundary. It doesn't enforce the stop. That takes a hard rule in the system running the work. It's an open standard: please contribute to the discussion! chrismdp.com/awml/

Sad that Sir Patrick Stewart has died :( I loved Picard on TNG as a teenager. This week I made my personal dashboard look like the bridge of the Enterprise, so I'm sharing it in his honour. I was tired of endless Telegram and Slack updates from my bots, so I decided to try putting everything back in one view to see how it works. What do you think?

I made this when METR said GPT-5.6 Sol exploited parts of its evaluation environment. We're way, way past that now. An agent can pass a test by changing the test. We must keep those checks outside the agent's reach. This is increasingly harder to do. metr.org/blog/2026-06-26-gpt…

Bests Opus 5.5 and cheaper per task on Artifical Analysis’s benchmark - so takes a spot on the Pareto Line. Nice one Demis 🙂

I think most AI at work is still powered by homemade generators. If it stops when one person is off, you need to build the sockets. That person knows which prompt works, which connector breaks and what to do when a task goes wrong. Everyone else depends on them. If this is how AI runs in your firm, give the champion help. Put the repeatable parts into a maintained service, and make it possible for colleagues to understand what the system did. A champion should help you build the next way of working, not spend every day holding up the current one! chrismdp.com/seven-ai-predic…

The new normal. chrismdp.com/comics/black-oc…

Credit to Anthropic - they don't hoard their own knowledge about how to use AI effectively.

Might have reached my cognitive limit in a working session last week. One Claude implementing AWML on my own system, mapping out two workflows at a time, and building a third into a new secure environment. Another Claude is working through the spec, building AWML workflows to help build... AWML - fastest ever software factory I've put together. The third is building out the AWML registry web app as another factory, trying frantically to keep up with the spec changes. I'm building a language for describing who does what in a workflow and now I can understand workflows better, Claude can work so much harder on them. The main limitation for my productivity is me, then the tools I have. As we build better tools, the sky's the limit (or my brain explodes). chrismdp.com/awml/

I don't think an AI rollout starts with permission to use the tool. It starts with changing the work around it. If every handoff and approval remains the same, faster drafts arrive at the same manager's desk. That manager's day fills up while the team is told it has been transformed. Pick one piece of work. Decide which decisions still need a person and which handoffs can go. chrismdp.com/ai-must-be-line…

I'm sorely tempted to make a game again. I used to build games professionally and as a hobby. Last year I tried using AI to make isometric assets for a gardening game. The pictures looked lovely. The tiles didn't line up, so I couldn't use them. That's the maddening thing about game development: an image can look right and still be wrong for the job. Scale, alignment and consistency matter more than a convincing screenshot. I'm tempted anyway.

I've been reading about the new Claude Duet - it's another place you might run your AI work. It's a good test for the idea behind AWML: describing the work separately from the platform. Say you want an AI to chase unpaid invoices. What starts the check? Which record holds the amount due? Who approves the message before it goes to the customer? What happens if the job stalls? Duet may be one way to do parts of that work. Its research preview can check connected apps, prepare a reply and ask before sending it. But the job belongs to your business, not to Duet. That's why I made AWML: to draw the steps, the people, the AI and the approval boundaries before choosing where they run. Duet goes in the platform view. It isn't the workflow. chrismdp.com/awml/

I don't trust AI workflows I can't draw, yet I had no way to draw one! I invented AWML to get around that. Through this process I spotted a prompt injection vulnerability in my sales funnel agent, and discovered I couldn't tell where my data was coming from in all cases. Amazingly helpful. Give AWML a try on your workflows, and give me feedback! chrismdp.com/awml/

I’ve open sourced a modelling language for agentic workflows (AWML) I’ve been working on. Modelling workflows independently of eg Cowork, Codex, Managed Agents etc gives you auditing, automatic optimisation, and intentional security boundaries. AWML shows who does each step, where information crosses permission boundaries, and which platform runs it. This is early stage so very keen to get feedback - comment or open a GitHub issue. GitHub: github.com/chrismdp/awml Blog: chrismdp.com/awml

A successful AI demo tells you very little about what's going on under the hood, and who's checking it. chrismdp.com/seven-ai-predic…

I use AI to help with writing and thinking, but I must stand by each word I put out. My new article explains where I draw that line. In short, don't be like Laura :) chrismdp.com/writing-with-ai…

Instead of selling books, will we be selling skills in the future, with the book coming free with the skill? chrismdp.com/slice-your-job-…

AI is good enough, but rollouts stall, and it's normally due to people problems. An example: put your best people in a small AI team and they'll move fast. The rest still have a product to maintain. They watch the new work race ahead while their own chance to learn gets postponed. You can connect the two: share what the team learns, without asking everyone to abandon their existing work. It sounds obvious, but you have to make sure it happens! This is one of seven people problems in my talk at Øredev in Malmö this November. If you're going, come and tell me how your team is handling it! oredev.org/program/25377809-…