Automation

The build was performed by an AI agent driving a real browser through Model Context Protocol tools. No Copilot Studio SDK, no API scripting outside Phase 3, and no recorded macro. This section documents how, and what breaks.

The tool surface

Tool family	Role
playwright-browser_*	Every UI action in every phase
filesystem	Writing captures and artifacts to disk
Azure CLI	Phase 3 only
Azure AI Search REST	Phase 3 only
@darbotlabs/darbot-browser-mcp	Independent validation of the audit report

The three rules

UI only. If a maker cannot do it in the browser, it is not part of the runbook. Phase 3 breaks this rule because Microsoft IQ has no creation UI at all, and that break is called out explicitly rather than hidden.

Screenshot after every action. Not after every step, and not after every milestone. This is the expensive rule and the one that paid for itself: the silent HTTP 500s in Phase 6 and the isConsentProvidedToChangeACPToAny=false flag were only recoverable because the network trace was captured alongside each UI state.

Self-reflection between steps. Read the resulting state before issuing the next command. The build was not a pre-planned script being replayed; each action was chosen after observing the previous result. This is why the runbook contains rejections and retries rather than a clean path.

What is fragile

Five failure modes recurred often enough to be worth naming.

1. Element references go stale on every snapshot. Use role and text locators, never stored refs.
2. Adjacent React fields jumble under type and fill. Use the native value setter through browser_evaluate.
3. Deep SPA navigation can drop the session and force re-authentication. In-app clicks are more reliable than browser_navigate to a deep route.
4. Clicking an agent name opens a new tab at index 1. Subsequent actions target the wrong page unless you follow it.
5. Canvas state accumulates. Capturing many configuration panels requires a deterministic reset between each one.

Each of these is documented with its workaround in Capture patterns.

In this section

• Playwright MCP — the tool inventory and the control-identification strategy.
• Capture patterns — the five failure modes and their workarounds.
• Workflow node capture — the deterministic reset loop that captured thirteen node types.
• Darbot Browser MCP — the independent validation pass and its known driver limitation.
• Scripts — the PowerShell and Python in the repository, including the one that is now a no-op.