QC'ing a New Agent in Assist 2.0

Who this is for: Anyone asked to check a new Agent before client use.

Goal: Use the Agent 10–20 times across real scenarios and log any issues ("snags") on the QC Monday board.

1) Before you start (2 mins)

• Open Assist 2.0 and start a new chat. Describe what you want to do in natural language — Assist will find and use the Agent for you, or you can pick it directly from your installed agents.
• Confirm you're in the right workspace (agents and knowledge are workspace-scoped, so a misplaced chat will pull the wrong context).
• Jot down 3–4 real scenarios you actually handle (e.g., first outreach, discovery call, pricing objection, troubleshooting).
• Open the QC Monday board in another tab.

2) What "good" looks like

• Follows your instructions (tone, length, format, channel).
• Stays consistent with the persona or role you set; doesn't drift.
• Gives realistic pushback/asks sensible questions; not generic.
• Handles constraints (e.g., "max 120 words", "bullets only").
• If asked to step out of character, gives clear, brief feedback, then returns to character when told.
• Uses Plan Mode sensibly when the task is multi-step — proposes a plan you can revise before it runs.

3) Run your tests (target: 10–20 uses)

Keep each run short (1–3 minutes). Mix these:

A. Setup & persona

• Define role, industry, size, pain points, style, and authority level.
• Switch persona mid-session to check reset behaviour.

B. Scenarios & channels

• First cold email (100–150 words).
• LinkedIn DM (3 messages each).
• Discovery call (5–8 questions; time-pressed).
• Negotiation (pricing, scope, legal/compliance).
• "Bad brief" test: vague or missing info — does it ask clarifying questions?

C. Difficulty & format

• "Be more sceptical" / "warmer but cautious".
• "Max 120 words" / "bullet points only" / "include 1 question".
• Add a curveball (new stakeholder, legacy system, budget freeze).

D. Feedback toggle

• "Step out of character and critique my last two messages (clarity, value, CTA). Then return to character."

E. Plan vs Execute (new in 2.0)

• For multi-step tasks, try once in Plan Mode (review and revise the plan) and once in Execute Mode (immediate run). Note whether the plan it proposes is sensible.

Tip: Note pass/fail quickly. You'll log one snag per issue you find.

4) Log snags on the QC Monday board

Use the standard columns:

• Name — short title of the issue
• Link to Output/Chat — paste the exact chat URL
• What task were you trying to do? — 1–2 lines of the scenario
• What happened? — actual behaviour (what you saw)
• What was the expectation? — 1–2 lines on what should have happened
• Severity — Low / Medium / High / Critical
• Screenshots (Optional) — attach or link
• Status (For Assist use) — leave blank

Severity guide

• Critical: breaks core behaviour or blocks use.
• High: major inconsistency; frequent.
• Medium: intermittent quirks; workarounds exist.
• Low: polish (phrasing, minor formatting).

Example snag (Agent)

• Name: Breaks character without being asked
• Link to Output/Chat: https://…/assist/chat/abc123
• What task were you trying to do? Pricing negotiation with a Procurement persona (formal, concise).
• What happened? After my third message, the Agent started coaching me (not in character).
• Expectation: Stay in persona until I explicitly ask to step out of character.
• Severity: Medium
• Screenshots: (attached)

5) Finishing up

If you hit no snags, ping us via email to let us know QC is complete.

Need help? we'll jump on a quick walkthrough.