Agent Delegation

  • Desktop app
  • All plans

What it does

A developer agent that has just finished a feature does not test it itself: it calls one tool, run_qa_test, on the AgentsRoom Test Runner MCP server. AgentsRoom then spawns a short-lived QA agent, usually on a cheaper model, which drives the embedded browser, runs the scenario, returns a verdict (pass, fail or inconclusive with a summary, an optional screenshot path and optional console logs) and is destroyed. The developer agent reads the short verdict and moves on; screenshots and DOM dumps never enter its context.

The constraint is mechanical: the browser tools are removed from the developer agent at spawn time, so the only way it can test in a browser is to delegate.

Video (1:27): Agent delegation, shown in the real app. The project's test agent is switched from Sonnet to Haiku in the Tests panel, a Claude Opus agent makes a shop's Buy buttons work and delegates the browser test, a QA agent on Haiku appears under it, clicks Buy in the built-in browser, and the verdict "QA Bot: passed" comes back to the Opus agent.

Where to find it

There is no button. Delegation happens from inside an agent's conversation: the developer agent calls the tool when it needs a browser test. The QA agent appears as a temporary agent in the project's agent list, indented right under the agent that called it ("↳ QA · for <agent>", role QA), and the app switches to its console and its browser screen while it works; it disappears once the verdict is sent. The model of the QA agent is set per project in the right panel's Tests tab: link Configure a cheaper model for testing at the bottom, then Override for this project, the Model card (the picker shows a price badge, $$$ for Opus, $ for Haiku), pick one and Save defaults; the card then reads "Project override" and "QA Engineer · haiku" (Use global removes it). Without an override the global default applies (QA Engineer on Sonnet). The project's QA context (datasets, URLs, instructions the QA agent reads) is edited from the QA agent's QA context modal (tabs Datasets, URLs, Instructions).

How to use it

  1. Optional, once per project: in Tests, Configure a cheaper model for testing, override the QA Agent with a cheaper model (Haiku).
  2. Give a developer agent a task that ends with an explicit delegation ("make the Buy buttons work, then delegate the browser test to a QA agent: open localhost:4321, click Buy, check the menu shows Cart (1)").
  3. When the agent calls run_qa_test, AgentsRoom asks for your approval first unless you turned that prompt off (see Settings).
  4. The QA agent starts, opens the Browser tab, runs the scenario and reports back. Watch it in its own tab if you want.
  5. The developer agent receives the verdict and continues. On failure, the QA agent can also file a backlog ticket carrying the long-form details (scenario, screenshot, console logs).

To make QA runs reliable, fill the QA context once per project: named datasets ("admin user" with email and password), pinned URLs (login page, staging), and free-form instructions prepended to every QA session.

Settings

  • askBeforeTestTools (global, Settings; project override in Project settings): ask before an agent drives a test or automation MCP tool, including the QA test runner. A single agent can be exempted.

Agent tools (MCP)

None on the AgentsRoom-MCP server. The delegation call is run_qa_test on the separate AgentsRoom Test Runner MCP server, and the QA agent answers through submit_verdict on the AgentsRoom QA Tester server. The QA agent also gets the backlog_* tools to file a ticket on failure.

Providers

Cross-provider: the developer and the QA agent can run on different CLIs and models. A common split is a large model on the developer side (Claude Opus, Codex) and a small one on the QA side (Claude Haiku, Codex mini). The QA agent needs a provider that can use the AgentsRoom browser MCP tools.

Mobile

Not on the mobile companion. The QA agent shows up in the agent list like any temporary agent while it runs, and its verdict lands in the developer agent's conversation, which the phone can read.

Limits

  • Ask for the delegation in so many words ("delegate the browser test to a QA agent"). Filmed on 2026-11-10: with "then have it tested in the browser", a Claude Code agent tried to test by itself (Chrome tools, then any local browser) instead of calling the QA agent.
  • Browser tests only today. Electron app tests (via an AgentsRoom Electron MCP library) and React Native tests are on the roadmap.
  • Everything runs on your machine: the developer agent, the QA agent, the MCP bridge and the browser. Only the model calls go to the provider.
  • The QA agent tests the developer's checkout: when the developer works in a worktree, the QA agent inherits that worktree rather than getting a fresh one, so it sees uncommitted work.
  • One run is capped at 10 minutes by default. The timeoutMs argument of run_qa_test can only raise that cap: a shorter value is ignored, because a QA agent needs the default just to boot its CLI, connect the browser and read the scenario. A run that hits the cap comes back with the status timeout and a summary that says whether the QA agent had started (it started but never submitted a verdict) or never finished starting (no verdict was ever possible).

Common questions

  • How do I make my agent test its work on a cheaper model? Set the project's QA Agent to a cheaper model once (Tests tab, Configure a cheaper model for testing, Override for this project, Model, Haiku, Save defaults), then end your task with "delegate the browser test to a QA agent" and the scenario. A QA agent appears under yours, runs the test in the built-in browser and sends back pass or fail. Video: https://youtu.be/RkbcJaJrIf4
  • Is there a video that shows it? Yes, a minute and a half, filmed in the real app: https://youtu.be/RkbcJaJrIf4
  • Why not let the developer agent test itself? Cost (a QA pass on a small model is roughly ten times cheaper than on the large one), context (the developer stays clean), reliability (a single-purpose tester clicks better than a multitasking developer).
  • Can the QA agent file a backlog ticket when a test fails? Yes. Besides the browser tools and its verdict tool, the QA agent has the same AgentsRoom tools as any agent (backlog, project memory). Ask for it in the scenario ("file a ticket if the login fails"): it files the ticket before sending its verdict and puts the ticket id in the verdict details. Before 2026-10-16 its own instructions listed only the browser and the verdict, so it could answer that it had no AgentsRoom tools.
  • Does this need a team graph? No. It is the Dev to QA flavour of a team exposed as a single tool call. For multi-step pipelines with conditions and loops, use Agent Teams.
  • Can I stop the approval prompt? Yes, turn off "ask before test tools" globally, per project, or exempt one agent.
  • The developer agent only gets "timed out" and never a verdict. Read the summary of the timeout: "before the QA Bot finished starting" means the QA agent never came up (check that its CLI is installed and signed in), "started but never submitted a verdict" means the scenario, the page or the browser kept it busy past the cap. Asking for a cap below 10 minutes used to end every run this way; those values are now ignored.