3.12 — Through-line workshop: "It acts on its own" 🔥

Goal: turn the prompt from Level 2 into an agent that acts, on your project. We deliberately keep it minimal — one agent, one or two tools, one clear loop.

🟢 In plain terms — an agent = your prompt + tools + a ReAct loop (Think → Act → Observe → Adjust). In no-code: a trigger → Airtable (memory) → AI that decides → action, orchestrated by Make or n8n.

🎯 The block below describes the agent to build for your project (chosen on the home page).

🧠 Your documents copilot

• The agent: doc base (Airtable or files) → search tool → grounded answer (your N2 prompt).
• ReAct loop: think “what info do I need” → act “search the docs” → observe → answer with citations.
• Test on 3 questions, including one outside the docs (the agent must say “I don’t know”). 👉 Deliverable: an assistant that answers about 5 of your documents.

📥 Email sorter

• The workflow (Make/n8n): incoming email trigger → AI classifies → action (draft / right folder / urgency alert).
• Memory: an Airtable table = log of processed messages.
• Test on 3 emails, including one that forces a correction (bad classification caught). 👉 Deliverable: a workflow that handles a test inbox.

🎯 Lead machine

• The pipeline: source → enrich → AI qualifies (N2 prompt) → outreach draft → scheduled follow-up.
• Memory: Airtable prospect base (status, last contact).
• Test on 3 prospects, including one the agent must cleanly rule out. 👉 Deliverable: a pipeline that qualifies and prepares outreach for 10 test prospects.

📄 Document extractor

• The pipeline: dropped document → extraction (your N2 prompt, structured output) → validation → Airtable write.
• Loop: if a field is missing, the agent flags it rather than inventing it.
• Test on 3 documents of different formats. 👉 Deliverable: a pipeline that fills a table from 5 invoices.

✍️ Content factory

• The team (multi-agent justified here): Writer (per-channel drafts) + Reviewer (brand voice), orchestrated from 1 pillar.
• Tools: Make/n8n to chain steps; Airtable = editorial calendar.
• Test: 1 pillar piece → 3 reviewed formats. 👉 Deliverable: 1 pillar turned into 3 formats ready to review.

🎯 Your own project

1. Pick ONE useful tool (web search, file reading, calculation…). Just one, to start.
2. Draw the ReAct loop for your case: which Thought triggers the Action? What does the agent observe? When does it stop?
3. Sketch the no-code version (3.9 model): trigger → Airtable → AI decides → action. Make or n8n? Which table?
4. Test on 3 cases, including one that forces the agent to self-correct. 👉 Deliverable: the sketch of your agent + one test that runs.

🧭 One agent first. Only move to a team (CrewAI: Strategist → Researcher → Writer → Reviewer) if the task truly justifies it. Before splitting: wouldn’t a single agent be enough?

For techies (optional): implement the ReAct loop in a notebook (notebooks/niveau-3-react.ipynb), or build the same agent in no-code in Make/n8n + Airtable.