ANTIGRAVITY LABJP
Articles/Integrations
Integrations/2026-05-29Intermediate

Copilot Studio's Computer-Using Agent Next to How I Run Claude in Chrome Every Morning

Microsoft made Copilot Studio's Computer-Using Agent generally available. Here is how that announcement looks from the perspective of an indie developer who already runs Claude in Chrome against AdMob and Crashlytics every morning.

Copilot StudioComputer-Using AgentClaude in ChromeAdMob12CrashlyticsIndie Developer3AI Agents14

I have been shipping iOS and Android apps as an indie developer since 2014. The wallpaper apps I maintain — Dolice's lineup — have crossed 50 million cumulative downloads, and I keep the lights on by myself. On May 28, 2026, Microsoft moved the Copilot Studio Computer-Using Agent into general availability. The timing is interesting for me because I already use Claude in Chrome to check AdMob and Crashlytics every morning, and the two products are looking at the same problem from very different ends.

What just became generally available

Per Microsoft's announcement, the Copilot Studio Computer-Using Agent lets organizations build agents that interact with websites and desktop apps directly through the UI. The pitch is for systems that have no underlying API, where teams previously relied on brittle scripts or manual workarounds.

A few items in the GA bundle stood out from an operations point of view:

  • Safer credential management built into the platform.
  • "Adaptive" automation that does not break when a screen or page changes.
  • Computer-using steps can be embedded in multi-step workflows alongside API-based actions and approvals.
  • A new orchestration layer reportedly delivers about 20% better evaluation performance and 50% lower token consumption.

The "does not break when the UI changes" point is the one I keep rereading. As someone who automates AdMob and Crashlytics dashboards from a browser, that is the single most painful failure mode. AdMob redesigns its dashboard meaningfully about once every six months, and I have rewritten scripts more than once because of it.

What I have Claude in Chrome doing

Every morning, Claude in Chrome handles roughly this routine for me:

  • Read out yesterday's revenue, eCPM, and fill rate across six apps on AdMob.
  • Surface every new crash on Crashlytics that pushed Crash-free users below 99.7%.
  • Confirm the new mediation adapters — Liftoff, Unity Ads, InMobi — are still enabled on AdMob's UI.
  • Pull yesterday's installs by region from App Store Connect and Google Play Console.

These are all dashboards that either have no public API or have an API that does not cover what an indie developer actually needs. AdMob has the AdMob API, but the mediation waterfall is configured through the UI. Crashlytics requires the UI for crash filtering and stack-trace browsing in practice.

Copilot Studio's Computer-Using Agent and Claude in Chrome are looking at the same surface area, just from opposite scales.

Where each one belongs

Copilot Studio is aimed at enterprise business processes built with low-code and no-code tooling, with multiple agents and workflows composed together. The Work IQ interoperability layer is explicitly designed for agents, workflows, APIs, and tools to play in the same orchestrated environment.

Claude in Chrome is built to ride along with a single browser session and run a single dashboard at a conversational pace. You ask for something, it clicks and types in front of you.

In my head the split looks like this:

AxisCopilot Studio Computer-Using AgentClaude in Chrome
Target scaleEnterprise business automationIndie developer's daily operations
Credential managementEnterprise-grade, built inRides on the Chrome session
AdaptivenessPlatform-level handling for UI changesI rewrite the prompt to keep up
IntegrationsSits alongside APIs, approvals, business logicMostly one-shot browser actions
Failure behaviorCan fall back to an approval step in the workflowStops in the chat; I intervene by hand

If all I want is to run a single dashboard chore in five minutes every morning, Claude in Chrome is enough. If the work is "invoice → ERP entry → approval → customer notification" with human approvals between steps, Copilot Studio with Workflows is the better-shaped tool.

How to keep a UI automation alive when the UI changes

The headline new capability for Copilot Studio is automation that does not break when the screen changes. As someone touching AdMob and Crashlytics every day, this is the most important question.

To approximate it with Claude in Chrome, I have settled on a few habits:

  • Anchor on visible text, not selectors. I instruct against text labels like "Show ad units" or screen-reader-friendly attributes. I avoid CSS selectors and coordinates unless absolutely necessary.
  • Always confirm the page after a transition. After a click, I ask Claude to tell me where it is. If the answer is unexpected, we walk it back a step.
  • Keep steps short. Five steps per prompt is too many; recovering from a mid-prompt failure is painful. I break things into three-step units.
  • Separate reads from writes. "Show me yesterday's revenue" is a read; "Edit the mediation group" is launched as a different task. The read prompts run autonomously; the write prompts have explicit confirmation.

Copilot Studio appears to be absorbing all of this at the platform layer. The 20% evaluation-performance improvement in the new orchestration layer is probably, in part, exactly this kind of UI-change tolerance moving from the human prompt into the platform.

Lessons from six months of indie UI automation

Six months of running Claude in Chrome against AdMob has shaken loose a few gotchas worth writing down.

The first is credentials. AdMob operations need a Google session, so Claude in Chrome rides on whichever profile I am signed into. I now use a dedicated Chrome profile for AdMob with Claude in Chrome installed there, separate from my main profile. Conceptually this is the same problem Copilot Studio addressed with enterprise credential management.

The second is rate limits. AdMob's dashboard slows down under bursts of queries, and if Claude in Chrome retries during that window, the session can trip anti-abuse heuristics. I now require at least three seconds of delay between operations directly in the prompt.

The third is verification. If Claude in Chrome reports that it added Liftoff to a mediation group, I still open a separate tab and check by eye that it actually happened. This is the equivalent of Copilot Studio's approval-step gate. For an indie developer, I have decided not to skip it.

Real-time voice and where it is heading

This release also generally availabled real-time voice agents through Dynamics 365 Contact Center in North America. Caller identification, question handling, handoff to a human operator — all in voice.

I have not yet integrated voice agents into my indie ops. That said, I once spent a single session writing 72 review replies across four apps and six major countries on the App Store alone, and it occurred to me that text → voice → translation might be worth experimenting with there.

Microsoft pushing voice general availability in the contact center context tells me enterprise expectations for support experiences are shifting. Where indie developers should plug voice into review replies and inbound user mail is a question worth thinking about now rather than later.

A working selection guide

The deciding factors between Copilot Studio and Claude in Chrome are basically the size of the work and whether approvals belong inside the workflow. My personal heuristics:

  1. Approvals or audit logs required? If yes, Copilot Studio with Workflows. If no, Claude in Chrome.
  2. Multiple humans running the same workflow? Shared, Copilot Studio. Solo, Claude in Chrome.
  3. Central orchestration across agents needed? If you want to see agent nodes and approvals in one canvas, Copilot Studio. If single-shot dashboard ops are enough, Claude in Chrome.

I will keep Claude in Chrome as my main driver for the foreseeable future. The day Stripe membership operations, Apple, and Google notification triage cross a threshold where they need to be orchestrated together, that is when I will look at Copilot Studio seriously.

Putting them side by side

The Copilot Studio Computer-Using Agent GA is a big enterprise milestone. For an indie developer it is also a tailwind: "let the AI operate the dashboards that have no API" is being taken seriously at industry scale.

What I have been doing every morning with Claude in Chrome is, on a smaller scale, the same problem set. UI tolerance, credential management, verification, human gates — every one of those concerns shows up identically at both ends.

Starting tomorrow I am adding Liftoff, Unity Ads, and InMobi to the AdMob mediation. I will be running Claude in Chrome through it and watching how the Copilot Studio side evolves in parallel, picking up what fits my operation.

Share

Thank You for Reading

Antigravity Lab is ad-free, supported entirely by members like you. We publish practical guides daily with implementation code, benchmarks, and production-ready patterns. If you've found it useful, we'd love to have you on board.

  • Copy-paste ready implementation code
  • New advanced guides published daily
  • $5/mo or $10 for lifetime access
View Membership →

If you found this article helpful, a small tip ($1.50) would mean a lot to us. Your support helps keep this site ad-free and covers server and hosting costs.

Related Articles

Integrations2026-06-09
Rebuilding Wallpaper Image Delivery Around Resolution Buckets — Letting an Antigravity Agent Own Conversion and Validation
Every new device resolution quietly makes a wallpaper app heavier. I stopped shipping one master image to every device and rebuilt delivery around resolution buckets, WebP/AVIF, and an edge redirect — then handed conversion and validation to an Antigravity agent. Real code and thresholds included.
Integrations2026-05-30
Diagnosing AdMob Mediation Fill Rate with an Antigravity Agent — A Memo on the Revenue I Lost Chasing eCPM Alone
When AdMob mediation revenue stalls, the cause is usually fill rate, not eCPM. This memo walks through a diagnostic pipeline that lets an Antigravity agent read your BigQuery export and Firebase metrics to isolate per-network fill rate drops automatically, with real numbers from indie development.
Integrations2026-05-24
Six Weeks of Letting an Antigravity Background Agent Watch the Drift Between AdMob and Remote Config
A six-week field report on handing weekly AdMob and Firebase Remote Config diffs to an Antigravity Background Agent, with the drift cases it actually caught and the operational boundaries I ended up drawing.
📚RECOMMENDED BOOKS
Build a Large Language Model (From Scratch)
Sebastian Raschka
LLM Dev
Prompt Engineering for LLMs
Berryman & Ziegler
Prompting
AI Engineering
Chip Huyen
AI Eng
* Contains affiliate links
See all →