Routine-Powered Performance — Building Structured, Multi-Step AI Agents for Marketing
31 Jul 2025 • Facilitator: Performics Labs Newsroom
💡 TL;DR — Routine shows that explicit JSON plans + clean parameter passing slash tool-call errors and unlock enterprise-grade LLM agents. Use this guide to turn the paper into a working marketing framework.
🗺️ How to Navigate This Page
- Watch the 3-min explainer (video at the top).
- Pre-read the PDF — even a skim helps: Routine research paper.
- Open the interactive Notebook → https://notebooklm.google.com/notebook/09062800-9d65-4a44-bf03-3f29b2e6eb38 to explore highlights, Q&A, and key mind maps.
- Dive deeper with our contextual articles:
- News Brief — Routine in 7 Bullets
- Deep Analysis — Building Agentic Marketing Pipelines
- Join the discussion — add questions or vote on build ideas via the comment widget at the bottom.
- Code the challenge — scroll to Build-In-Public Sprint and pick a task.
🛠️ Build-In-Public Sprint
Goal: ship a Routine-style agent that auto-pauses poor-ROAS ad sets on Meta in under 14 days.
| Step | What to Deliver | Hints |
|---|---|---|
| 1 | Fork the starter repo | Uses Mastra TS + LangGraph |
| 2 | Draft JSON plan (max 5 steps) | See Figure 3 in the paper |
| 3 | Implement Executor & Meta Marketing API calls | Re-use our typed wrappers |
| 4 | Log state→action→reward tuples | Ray RLlib buffer template included |
| 5 | Record a 90-sec Loom demo & share in comments | Top 3 voted demos get featured |
🏁 Ready to level-up?
Join the Routine × Mastra Hackathon → · Register on Luma ↗
Timeline
• Submit by 8 Aug • Demo Day 11 Aug
Starter templates & helper code already in the repo — just fork and build!
🤔 Discussion Prompts
- How might Routine’s plan/execution split improve auditability versus classic prompt-chains?
- Would you distill to a 7B model or keep GPT-4o as Planner? Trade-offs?
- Which KPIs make best reward signals across Search / DV360 / AMC?
📚 Further Reading
🚀 Ready? Grab the repo, join the chat, and let’s build the next generation of performance agents — together.
Discussion & Idea Voting
Up-vote next week’s build idea by reacting with 👍 to any comment.
Published on Thursday, July 31, 2025