Updating an agent

Edit personality, swap models, add a tool, raise prices. Every change re-runs the eval suite — your live agent never silently regresses.

Three ways to update

1. Through Claude Code (fastest)

Tell Claude Code what you want changed. The MCP tool is update_website_chatbot. It applies the diff, re-runs evals, and reports whether the agent still passes the gate.

> Update Acme Dental's chatbot. Raise the cleaning price
  from $120 to $135. Add a "before-and-after photos" FAQ
  pointing to /gallery. Re-run evals.

2. Through the dashboard

Open Agents → pick the agent → Settings tab. Change Soul, Brain, or tools. Hit save. The agent moves to draft status and you have to re-run evals before it can go back to live.

3. Edit the skill-pack directly

Agent intelligence lives in markdown skill-packs at packages/crm/src/lib/agents/skills/. Edit the markdown, commit, deploy — all agents using that skill-pack get the new behavior. Power-user move; useful for agencies updating 100 client agents at once.

What happens to live agents during an update

Zero-downtime by design

While you're editing, the live agent keeps serving traffic with its previous version. The new version is in draft until evals pass and you publish. There is never a window where a half-configured agent is talking to your customers.

Common updates

  • Change a price. Edit the FAQ snippet → re-run evals → publish. The old conversation history stays attached to the agent.
  • Swap the model. Settings → Brain → pick a new model. Re-run evals (different models sometimes change pass rates by 5–10%).
  • Add a tool. Settings → Tools → enable e.g. send_followup_email. Add a scenario testing the new tool. Re-run.
  • Tighten a refusal rule. Edit the Soul's refusal list. Add a scenario where a customer pushes on the rule. Confirm the bot holds.

Versioning and rollback

Each publish creates a versioned snapshot. If a v3 update tanks eval scores or starts misbehaving in production, hit "Roll back to v2" in Settings → Versions. The previous version goes live instantly; v3 stays in draft for you to debug.

Next