Skip to main content
🚫 Most customer support AI agents are reactive. No context. No memory. No idea what users actually need next. ✅ A proactive onboarding agent is different. It watches, learns, and acts — surfacing the right next step before users have to ask.
📡Real-time event tracking
🧠LLM context assembly
💬Proactive chat + visual triggers
🗂️User memory & workflow ontology
This is the blueprint for going from a reactive support AI agent to a proactive onboarding and adoption agent. Walkthrough — same recipe in video form. Open on YouTube if the player does not load.

​
The building blocks

Follow the steps in order — each page stands alone so you can ship incrementally.
  1. Connect real-time actions — Stream structured ActionsPayload events into your stack.
  2. Add actions to LLM context — Combine live actions, chat history, and optional KB into one inference-ready context.
  3. Define proactive triggers — Define when to interrupt, with Agent State v2 gating so proactive delivery stays non-intrusive and non-overlapping.
  4. Connect relevant visual guidance — Route qualifying proactive outcomes into on-screen tours (chat vs tour vs silence) using your delivery channel.
  5. Enrich with user memory — Define workflows so Atlas and memory expose completion rates and mastery; then filter with cross-session user memory.

​
Implementation references

Use these SDK references as you wire the recipe into production:
  • UserSessionIndex — required for production user-keyed chat so user_id resolves to the correct recent sessions before Step 2 context assembly.
  • compose_chat_pipeline(…) — compose safe default chat-ingestion wiring in Step 2.
  • build_copilot_app(…) — stand up a minimal FastAPI bridge when you want an out-of-the-box serving layer.
You can build and ship real-time event ingestion today. Visual guidance wiring is covered in Step 4. Workflow ontology + user memory enrichment (Chrome extension; docs with Autoplay via Slack) is part of Step 5; expect an iterative loop on Autoplay Atlas labelling quality. The in-SDK knowledge base query and full user memory rollout are still rolling out.