Raybot

Raybot is a Ray and Ray-clusters expert agent, built end to end inside Microsoft Copilot Studio's next-generation runtime — codename Coral — using nothing but browser automation.

This site is the complete record of that build: the runbook that produced the agent, the ground-truth UI and API audit of the runtime it was built in, a typed object model and SDK derived from both, the recovered terminal history, and all 95 screenshots.

What makes this different

Most agent documentation describes what a product is supposed to do. This one describes what it actually did, on a specific tenant, on specific dates, with a screenshot for every assertion.

• Every step is UI-only. The agent was created, configured, published, and tested through Playwright MCP driving a real browser. The one exception is Phase 3, which needs Azure CLI and the Azure AI Search REST surface because no UI exists for it yet.
• Every step was captured. A screenshot was taken after every action, not after every milestone. That discipline is what surfaced the silent HTTP 500s and the tenant consent flag that the maker UI never explains.
• Every claim carries a confidence level. The audit distinguishes CONFIRMED (observed live) from SOURCE-INSPECTED (read from code) and UNCONFIRMED (inferred). This audit is CONFIRMED throughout.
• The failures are documented as carefully as the successes. Raybot's Teams channel never reached the tenant catalog and its evaluation scored zero percent. Both are explained, with the network evidence that proves neither is an agent-quality problem.

The agent

Property	Value
Agent ID	4de4ada9-ad61-4c22-a0d0-a0bd1c189f4e
Schema name	Default_Raybot_vyN1Rz
Environment	Cypherdyne, Default-6b104499-c49f-45dc-b3a2-df95efd6eeb4
Model	Claude Sonnet 4.6
Knowledge	https://docs.ray.io
Skill	ray-cluster-troubleshooter
Microsoft IQ	Foundry IQ knowledge base raybot-kb
Memory	Enabled
Built by	darbot@timelarp.com
Re-audited by	dayour@microsoft.com

The audit

Measure	Count
Surfaces audited	16
Elements catalogued	1,627
Interactive elements	1,115
Elements carrying a test id	983
Unique test ids	177
Test ids shared across surfaces	76
Interactive elements with no derivable label	228
Network requests observed	1,249
Functional requests classified	611
API endpoints mapped	57
Screenshots	95

Where to start

• If you want to reproduce the build, start at the Runbook.
• If you want to understand the runtime, start at the Coral schema.
• If you want to write code against it, start at the Object model.
• If you want to see it, start at the Gallery.
• If you want to know how this record was reconstructed, start at Session forensics.

Honesty notes

These qualifications apply to everything on this site.

1. isInteractive in the element audit is computed from native control type, href, ARIA role, tabindex, and contenteditable. It is independent of whether an element carries a data-testid. An element can be interactive with no test id, and can carry a test id while being inert.
2. derivedLabel is a heuristic, not a strict accessible-name computation. The heuristic used for each row is recorded in labelSource. Treat the 228 label gaps as a lower bound on accessibility work, not as an audit finding.
3. The API map is observed network activity, not a published contract. Endpoint names, shapes, and behaviours were read off the wire. They can change without notice.
4. This record covers a preview runtime on one tenant. Tenant policy materially changed outcomes: the Teams blocker is a tenant consent gate, and the Microsoft IQ blocker is an Azure SKU gate. Neither is universal.
5. There are no emojis anywhere on this site, by convention. Status is conveyed by words and badges.