cgray-redhat Agent Use
Classify the user's question, route to the correct agentic interface, and hand off with a crafted prompt — or delegate to an existing Cursor skill when MCP is the better interface.
Interaction model: Route + hand off only. Do not automate web agent UIs with Playwright unless the user explicitly requests browser automation.
Prerequisites
- Run refresh-connections if Dataverse MCP delegation is likely.
- Read project
dataaccess.md(if present) when routing data questions to MCP. - When routing is ambiguous or docs may be stale, refresh Source context:
python3 ~/.cursor/skills/agent-use/scripts/fetch_agent_docs.py --format json
If the script fails with auth errors, run the thesource skill authentication flow first.
Workflow
Agent Use:
- [ ] Step 0: Intake — clarify intent and deliverable
- [ ] Step 1: Classify — sales | enterprise_data | people | access | docs
- [ ] Step 2: Route — hand off to web agent OR delegate to Cursor skill
- [ ] Step 3: Report — rationale, URL, prompt, sources
Step 0 — Intake
Use AskQuestion when available; otherwise ask in one message.
| Question | Purpose |
|---|---|
| What are you trying to accomplish? | Core routing signal |
| Is this about selling/deals/RHSC or enterprise data/analytics? | Primary fork |
| Do you need the answer here in Cursor, or is a web agent chat OK? | MCP vs web agent |
| Any scope filters? (account, geo, time period, person) | Crafted prompt |
Gate: Do not route until intent is at least sales, enterprise_data, people, access, or docs.
Step 1 — Classify
people lookup (who is X, title, manager)?
→ dataverse-people-lookup
sales / RHSC / quote / TDP / account plan / competitive / meeting prep?
→ Sales Assistant handoff
enterprise data / pipeline / renewals / bookings / revenue / analytics?
├─ user wants answer in Cursor OR needs access check?
│ → guided-data-analysis
└─ user wants conversational exploration with charts?
→ Dataverse Agent handoff
access / permissions / "do we have data for X"?
→ guided-data-analysis (Steps 1–3)
"what can agent X do?" / capability question?
→ fetch_agent_docs.py or thesource + explain + handoff
See reference.md for the full routing matrix and example prompts.
Step 2 — Execute or hand off
Web agent handoff (Sales Assistant or Dataverse Agent)
Present using this template:
## Recommended agent: [Sales Assistant | Dataverse Agent]
**Why:** [1–2 sentences tied to Source capabilities]
**Open:** [URL]
**Suggested prompt:**
> [Crafted question with scope, account/geo/time if known]
**Tips:** [workflow library, Memories, Favorites, or starter question]
**Support:** [Slack channel]
| Agent | URL |
|---|---|
| Sales Assistant | https://shadowbot.io.redhat.com/agents/sales-assistant |
| Dataverse Agent | https://agent.dataverse.redhat.com/ |
MCP delegation (stay in Cursor)
| Intent | Delegate to |
|---|---|
| Employee title, location, manager | dataverse-people-lookup |
| Enterprise data analysis, access checks, SQL | guided-data-analysis |
| Auth / connection issues | refresh-connections |
Invoke the target skill; do not duplicate its pipeline.
Step 3 — Report
Always include:
- Which interface was chosen and why
- What was executed in Cursor vs handed off
- Source pages consulted (or
reference.mdif Source fetch was skipped) - Next step if access denied or the wrong agent was chosen
Quality rules
- Never fabricate agent capabilities; trace to Source excerpts or reference.md
- Prefer MCP in Cursor for governed data when permission probing matters
- Prefer Sales Assistant for seller workflows; Dataverse Agent for conversational data exploration
- Ask a clarifying question before handoff when intent is ambiguous
- Do not use Playwright for web agents unless the user explicitly requests it
Example prompts
- "Help me build a quote for Acme Corp with RHEL and a 15% discount"
- "What's our renewal pipeline risk in Germany for Q3?"
- "Which agent should I use to prep for a customer meeting?"
- "What can Sales Assistant do vs Dataverse Agent?"
Additional resources
- reference.md — routing matrix, URLs, example prompts, MCP delegation map