Not just a tool. A team in a platform.

We opened the first issue of this newsletter saying Gamebeast was about to speak MCP. It goes live Monday, reading first, with the writes that let it act following over the coming weeks. Connecting an agent is the small part; the rest is built so one can work unsupervised, with access scoped tightly enough to hand over, analytics it can query directly, and experiments it can start and read the results of by itself. Here is what that will open up:

Plan, then build it. Ask what’s worth changing and it reads your live numbers, writes the implementation plan, then goes and implements it.
Test its own work. It can run an experiment on what it just shipped and come back to read the result, rather than guessing whether it helped.
Open the PR itself. When the numbers point at a fix in your code, it opens a pull request against your codebase, with the data that motivated it in the description.

Plan, build, test, ship, read the result. That is the whole loop, and it will run without you. Reading lands Monday and the rest follows over the coming weeks, so if there is something small you would hand to it first, have that ready.

Experiments v2 hits general availability Monday

Last mention, we promise. From Monday it is generally available to everyone. Several tests can run at once as long as they don’t overlap, edits reach a running experiment instead of stopping at a frozen copy, and targeting goes much further than it used to: custom properties, computed ones like days since first played, logical groups that nest, plus new assignment types and sticky assignments. The rebuilt breakdowns and filters are in too. The dedicated results view is still coming. If you held off while this was settling, it’s settled.

Read the docs

Query caching is live

We walked through this last issue while it was still wrapping up. It shipped on September 2 and is on for everyone, with nothing to enable. Boards you’ve already opened come back right away and refresh behind you, tiles tell you when they’re queued or running instead of spinning, and the heavy questions that used to get cut off at thirty seconds now run to completion.

If you shelved a question because the dashboard couldn’t hold it, it’s worth opening again.

Thank you for coming to the mixer

Thank you to everyone who came out to the mixer with Twin Atlas on day one of RDC, and to Twin Atlas for hosting it with us. A full room of Roblox creators, community members, and studio teams, and more good conversation than we could get through in three hours.

If we talked about something and you’re waiting on us to follow up, reply to this email and it’ll reach us. If you missed this one, it won’t be our last.

On the horizon

queued descriptions everywhere

Anything you can name, you’ll be able to describe: queries, dashboards, configs and individual config fields, markers and their arguments in data management, and more. Written by you, or written by your agent.
This pairs with the MCP story above. When an agent instruments something in your code, it can leave the explanation here too, so the reasoning lives next to the data instead of in a chat log someone closed. That context is there for the agent next time, and for the human who inherits the dashboard and wants to know what “retention_v3” was actually measuring.

queued dashboards you can actually lay out

You can already reorder the queries on a board; what you can’t do yet is shape it. That changes: resize a query so the important one is the big one, drop in markdown to write headings and context between charts, and give each dashboard its own icon and color so you can tell them apart at a glance. More to come beyond that. A dashboard should be able to explain itself to whoever opens it next.

Not just a tool. A team in a platform.