Tyler Brown

Heya, I'm Tyler 👋 a principal engineer at @Atlassian. I've been building here for a long time, and working on AI since the GPT-3.5 days, when you had to figure out what you really could do with a 4K context. good times (seriously! it was so fun). I'm going to start sharing more of what i build and how i work: agent workflows, experiments, and lessons from getting AI to do useful things. I'm still learning plenty myself since it's changing every day. When did y'all start building with AI?

Forget Navier-Stokes, can someone please set AI on the case of figuring out how car seat buckles seemingly get twisted the wrong direction at ease when no one is watching, yet are incredibly difficult to manually flip back the right way?

It's crazy to think we thought things were moving fast a few years ago because React was adding something new and we all had to learn it. And then in these last two years we've gone from something that could sort of write decent tests to something that literally can optimize an entire part of your request path / pipeline by just saying "please make this super duper fast"

A lot of people are talking about how fast the Claude sub runs out, meanwhile I'm still amazed every day how long it lasts as someone whos used to seeing how quickly budget can get used up at actual API prices at work. Granted my side projects are also a lot smaller, but still.

Still eagerly awaiting when we can get an always on agent that works at companies with the usual enterprise security policies and all that other fun stuff. 🤞

Before AI: I love problems where the goal is to make a number go up or go down. After AI: I love problems where the goal is to make a number go up or go down and now I can do a bunch of them in parallel. It truly has become like a video game.

Skills over MCP support cannot come soon enough. Eagerly awaiting it! We've been working around it at Atlassian when we really do have a big chunk of required guidance (eg how to create a page) to give an agent, by effectively making "skill as tools" but having an actual established pattern would be great to shift to. Optional but helpful skills will also be a huge unblock of course.

I know the anti MCP rhetoric has really slowed down, but just to throw more one data point in the ring, the MCP growth were seeing at @Atlassian the last month is crazy. We just published numbers saying we're getting 15 million calls a day, up from 5 million in July, but my secret sources tell me its already rocketed well past that. MCP really is having its time to shine. On top of all the cool stuff already released / in the pipeline. Ask me anything about the Atlassian MCP! Jira, Confluence, whatever, if i don't know, I'll find someone who does.

We added soooo much new Confluence stuff to the MCP. You can now work with whiteboards, databases, create remixes, and soooo much more. If you have any requests or questions just drop em here!

/show-me-in-confluence is a skill I use every day. Love having a shareable record of something. If it was a good enough explanation for me why not make it so others can see it as well?

At this point "with an army of luna subagents" really is the answer to a lot of questions. Evaling your MCP or CLI, checking a bunch of UI operations and grabbing screenshots with browser use, fanning out and doing research, etc. What else are ya'll doing with it?

And just like that before most Jev powered use cases have truly hit production, competition is heating up. Great to see! I haven’t seen this much hype around a “new” AI paradigm in years and it’s awesome. It’s so fun thinking through all the new experiences we can build.

The Atlassian experience is now in ChatGPT with the new MCP extension points from the great @mxstbr and team! A home view where you can get right to work on whatever ya need to get done with one click. But honestly the best part might be the Composer @ mentions. You can search for work items, pages, and users by name and attach it as context without ever having to leave the chat box. One of my favorite things I've gotten a chance to work on in a while because I'm gonna use it literally every day. Check it out! What else do ya wanna see?

Confluence has gotten faster, like, a lot faster! The view page now loads 2.4x faster than it did in March 2025. We've been working super hard at this because we know that when people are in the UI these days they don't want to be waiting around. Is there anything more satisfying as an engineer than working on a problem where you just get to keep iterating and making a number go down? What do ya wanna see us tackle in Confluence next?

I'm finding that although Jev isn't "more accurate" than other models, if you have a limited amount of time and tokens to spend on a request (like so many real world use cases!), you can get higher accuracy because you can do so many more runs with Jev.

I've been absolutely loving playing around with Jev, so many possibilities. Sharing a real world use case at Atlassian. The first case I had to try was recommending where a new page should go in a Confluence space based on it's content. Did some evals with 100 pages and their ai summaries of what they are about (we always have these pre-computed already so it's not just something for the evals) and then ran 400 tests. Each model got two iterations, a first pass, and then a prompt improvement pass. The results have me beyond excited. With Jev we'd be able to classify at a scale we didn't think we could offer before with traditional LLMs. Cannot wait to play around with this more. Next up - organizing entire spaces from messy into clean with Jev plugged in wherever we can. (I've sent Jev over 30 million tokens and spent less than $1.50 - loving that too)

Dogfooding the CLIs and MCP servers we're building has shown how tenacious agents are at working around problems, and how easy it is not to notice. The first time I ran into this, we were just getting started on the CLI and creating a page under a parent was broken. The agent figured out the separate move operation still worked, so it created the page first and moved it under the parent afterwards. I wouldn’t have spotted the bug from the finished page. I only saw it because I was reading the actual tool calls it was making. Another time, I wondered why my Confluence pages looked better than other people’s. My agent had dug through an internal repo for implementation details missing from our skill. We added those once we realized. Other users couldn’t access that repo, so they’d been missing instructions I didn’t even know my agent needed. With a UI, I’m the one dealing with the friction. Agents won’t stop until they solve the problem. So they try another operation, find extra docs, then save it to memory for next time. Daily use still matters, but it will lie to you if you only look at the final artifact. Just goes to show even more why isolated evals are so critical. Anyone else finding their agent is sneakily working around bugs they didn’t know they had?

What makes a products MCP feel "good" to use? Do you actually review the tools yourself to see what they are doing? Or do you just check if your agent says it's struggling? Are we still as worried as before about tokens now that harnesses have gotten so good at optimizing tool calls?