Research and engineering
An agent you let into your Mac should be measured on what people actually ask of it, scored by what it actually changed, and honest about what it still gets wrong. Here is how we build Brello for Mac, with the numbers.
Brello is an agent that changes real things on your Mac: it moves files, writes documents, changes settings and deletes what you tell it to. A chat app can be judged by how its answers read. An agent has to be judged by the state it leaves behind: which files moved, what was deleted, whether a setting really changed, and whether Undo put everything back. Every post here starts from that.
Three ideas run through all of it. Code first, model second: what code can read whole, such as a date, an app’s name, a setting or a folder, code decides, and Apple’s on-device model is asked only for language. Private by construction: one module may reach the network, a test reads every source file on every build to keep it that way, and every request is written down before it leaves. Ask first, then Undo: risk is decided by code, not by a model’s account of itself, and every change Brello PI makes can be reversed.
The numbers on these pages come from the product itself, played end to end on made-up Macs and on a same-shape copy of a lived-in one, never from a benchmark the product doesn’t run. Where something failed, the post says so and says what changed.
Start here
The system card is the one page to read if you read only one: what Brello for Mac does and doesn’t do, the evaluations it was held to, its known limits and what it never does with nobody at the Mac. How we test Brello explains the four layers of testing behind every number on this site.
- System cardBrello for Mac system card
What Brello for Mac does and doesn’t do, the evaluations it was held to and their results, its known limits, what it needs to run, and what it never does with nobody at the Mac.
- ResearchHow we test Brello
Four layers, each played on the real product: 310 first requests on both paths, 353 scenarios in 1,301 phrasings on a same-shape copy of a lived-in Mac, every page and control in the app itself, and 2,475 package tests.
Research: measured on real requests
What people type to an agent on their first day, how dates and money are read the way each region writes them, and how every change is undone. Each post names the suite it was measured with, how many cases, and what the first run got wrong.
- ResearchThe first ten minutes
We wrote down 166 things a new person types first, played each on the real product, and scored what it changed. The first run handled 76 well. The suite now holds 310 cases, and passes every one, on the command line and in the app.
- ResearchUndo everything
Risk is decided by code: five tiers, one question for the whole plan, deleted files moved to the Mac’s Trash, and an undo checked byte for byte. On 65 tasks played three times, every run that changed something was put back.
- ResearchReading time the way people write it
“10/16” is October 16 in the United States and nonsense in Britain. How Brello reads days, times and money in code, by the Mac’s region, and says what it can’t read instead of guessing.
Engineering: how it’s built
How a Mac app earns its place on a Mac: a Finder that finds files by what you remember, settings changed and read back from macOS, one click that balances the load across the processor’s cores, a health page that stopped being the busiest thing on the Mac, and the small fixes that make it feel native. Each post shows the measurement before and after.
- EngineeringBalancing the load on a Mac
One click moves your own heavy background work onto the efficiency cores and, only on your yes, quits apps you haven’t used. How it chooses, the public call that does it, why its first read-back lied, and what it never touches.
- EngineeringSmall things that make a Mac app belong on a Mac
Text that drew upside down, a click that waited on a model, a spinner that cost frames, and files that now fall into the Dock’s Trash. The fixes nobody lists, and how each is held by a test.
- EngineeringWhat a health page taught us about a busy Mac
The page that shows what’s busy was the busiest thing on the Mac. What we changed, why memory pressure beats memory used, what keeps a Mac awake, and why every sentence on it is written by code.
- EngineeringCode first, model second
Brello PI reads what code can read, such as apps, dates, settings and routing, and asks Apple’s on-device model only for language. What the model is trusted with, what it never is, and what that costs in time.
- EngineeringPrivate by construction
One module may reach the network, and a test reads every source file on every build to keep it so. Every request is written to a ledger before it leaves, and carries only the words you typed.
- EngineeringA Finder that understands you
Home finds files by sentence, proves duplicates byte for byte, and groups downloads by where they came from. And the work that took the window from 6.5% of the processor at rest to under 1%.
- EngineeringSettings by asking
Say “turn on dark mode” and Brello changes it, reads it back from macOS, and offers Undo. Which settings an app may change, which it may not, and how each change is proven.
Questions about a post, or a result you can’t reproduce? Write to hey@brello.ai. For what leaves your Mac and how to check it yourself, see the trust center.