Core Concepts
How the agent decides
The diagnostic agent reasons in plain English over your past tickets and live tool data — no black-box confidence score. It either replies, asks for info, escalates, or proposes a code fix.
No more confidence thresholds
Earlier versions of DuggAI gated handoffs on a 0.0–1.0 confidence score. We retired that in favor of explicit LLM reasoning over real precedent — every decision now ships with a hypothesis the agent wrote and the past tickets it pulled. Operators can read why, not just trust a number.
The four outcomes
On every customer message, the diagnostic agent picks one of four actions and writes its reasoning into the diagnostic packet:
| Action | When it fires | What you see |
|---|---|---|
| canned_reply | Question is answered by docs or past tickets in your team's voice. | Tier-1 auto-reply sends; conversation closes when the user doesn't come back. |
| request_info_from_user | Missing IDs, version numbers, or screenshots before the agent can investigate. | Bot asks the customer a clarifying question — no human in the loop yet. |
| escalate | Real bug, account-specific data access needed, or a write the agent isn't allowed to autonomously run. | Card lands in your Inbox with the hypothesis, tool calls, and a proposed fix attached. |
| spawn_cursor | The fix is code. Reproduces the bug, points at the file, and dispatches to your Cursor cloud agent. | Coding-run card with streaming diff and PR link — still needs your approve before it merges. |
How the agent reasons
- Reads your operating manual (voice + investigation pattern) — auto-extracted from your Discord at onboarding.
- Searches the unified knowledge base (resolved tickets + docs) via
search_knowledge_baseand reads the closest 10 past tickets. - Lists which connected apps it can pull data from (Stripe, Sentry, GitHub, Supabase…) and calls the tools it needs.
- Writes a hypothesis in plain English. Picks one of the four actions above. Writes the reply or proposed plan.
The review queue
For escalate and spawn_cursor, the agent doesn't guess — it surfaces the full reasoning to your team via the Inbox. Approve as-is, edit the plan, or reject with a reason. The diff between the agent's draft and your edit becomes implicit training data — it's indexed back into the knowledge base so the next similar ticket arrives better-drafted.
Why this is better than a confidence score
- Auditable. Every escalation comes with the hypothesis and the precedent the agent pulled. You read the why, not just a number.
- Self-correcting. The model improves as your team approves, edits, and rejects — not as you tune a threshold.
- Voice-aligned. The operating manual + past tickets ground the agent in how your team actually talks and investigates, not a generic LLM tone.