One master agent. Twelve coding agents. Six websites, built in parallel. This is Armadai: a native Mac app that puts Claude Code, Codex, your browser and your projects on one infinite canvas, with one agent running the rest. Here's everything it can do 👇

Codex reset happening at 12 pst and my claude resets tomorrow at 6am. Looks like I will have 2 full $200 dollar plans to use this week! (both are tapped out now... using grok while I wait)

6 new bots and routines on top of the ones I already run. funnel audit, pain-signal leads, sponsor sweep, the whole setup as a template.

For those that use claude code, codex, and grok... what do use cases do you find best for each? When do you decide what provider gets what task? What is the best task to give to what provider?

I emailed support and did get a response saying there are issues and not to do anything at this time. I will update this thread if / when it gets resolved.

Browser agent screenshots can be hard to match to actions. In Chrome, agent-browser's screenshot --annotate numbers interactive controls. Label [2] maps to @e2 in its commands, so the image and action use the same reference. github.com/vercel-labs/agent…

A coding agent can add an await and still write a flaky UI test. In Playwright, these checks do different jobs: expect(await locator.isVisible()).toBe(true); await expect(locator).toBeVisible(); The first reads visibility immediately, then checks the returned boolean. The await doesn't make isVisible() wait for the element to appear. The second retries the visibility check until it passes or the assertion timeout is reached. Playwright re-fetches the element as it checks. That difference matters when a message appears after a request finishes. Checking too early can fail even though the UI is still updating. For example, if your app exposes a save-status test ID, you could check the expected result with: await expect(page.getByTestId('save-status')).toHaveText('Saved'); Choose a condition that represents the result the user needs. "The button was clicked" and "the save confirmation appeared" are separate checks. A few details to keep in the review: • Use Playwright's expect from @playwright/test. • Await the retrying assertion itself. • Assertions have a 5-second timeout by default; set a suitable limit for the behavior you're checking. This is a docs-based example, not a test run. When an agent writes a UI test, check what each await is waiting for. playwright.dev/docs/best-pra… playwright.dev/docs/test-ass…

Sending an empty list to an MCP tool? FastMCP 4.1.4 fixes generated CLIs that dropped empty arrays. An explicit [] could otherwise silently use the server's nonempty default. Check scripts where [] and a missing argument mean different things. github.com/PrefectHQ/fastmcp…

When a coding agent changes your UI, how do you check that it still looks right?

Been working on a large rebuild for a client here is a message from my agent "The archive export has passed one million rows with no errors. After the remaining rows are exported, the restore audit will verify the backup before Live editing opens" Pretty insane what you can do with Ai!

An agent's test can pass on the wrong exception. pytest.raises(RuntimeError) also accepts NotImplementedError because it's a subclass. If the exact type matters, assert excinfo.type is RuntimeError too. docs.pytest.org/en/stable/ho…

Building search over your own docs? Ollama's /api/embed can turn a batch of texts into vectors. Set truncate:false to get an error when input exceeds the context window, instead of cutting it off. docs.ollama.com/api/embed

What makes you stop a coding agent and take over yourself?

I keep hearing developers say, “I don’t use a VM, I use a VPS.” A VPS is literally just a virtual machine sold as a hosting service (in most cases). VM describes the technology. VPS describes the service. You're not using something different. You're just renting it.

Review a browser agent's plan before it clicks. Stagehand v4's observe() returns suggested actions with selectors, methods and arguments. You can inspect them, then pass them to act() to execute. docs.stagehand.dev/v4/basics…

Running grok bot app on your mac uses around 3gb (correct me if I am wrong) but running grok bot inside your X chat uses nothing and can do the same thing. Run grok bot in your chat theres no reason not to.... happy to hear other opinions / thoughts on this.

Pulled the replies under a thread about why the same account swings from 20K impressions to 200. Biggest takeaway from ~40 people: replies beat solo posts right now. Full write-up in the article — whats working for your follows this week?

At this stage I have spend 10's of thousands of hours building. My next goal is to spend 10k hours learning to market, until I have mastered it.