How to build an AI agent in 90 seconds — real step-by-step

You've seen that chat widget in the bottom-right corner of websites — the one that answers questions about the product instantly, with sources. It looks expensive and complex to build. It isn't. This tutorial shows you how to create yours in about 90 seconds. No code required.

What you'll have at the end

An AI agent trained on your own content, embedded as a widget on any page, responding with chunk-level citations and a visible confidence score. Production-ready.

Estimated time: 45–90 seconds from zero to live widget.

---

Step 1 — Create your account (10 seconds)

Go to simple-agent.me and click Early access.

The onboarding asks for your email only. No credit card, no mandatory OAuth, no long form. A magic link arrives in 8–12 seconds.

Why magic link? Because 63% of SaaS onboarding abandonment happens on the signup screen. Less friction = more people reaching the "wow moment."

---

Step 2 — Name and configure your agent (15 seconds)

On the second screen, you set:

• Agent name — appears in the widget header (Support Bot, AskAna, etc.)
• Base URL of your site — the system automatically extracts your favicon and primary brand color

The agent avatar is generated from your favicon. You can swap it later, but most of the time it looks great out of the box.

---

Step 3 — Add a training source (20 seconds)

This is what turns a regular chatbot into an AI agent that actually knows your business.

Three options:

1. URL — Paste any public URL. The system crawls the page + internal links, up to 100 pages.
2. PDF — Direct upload. Works with manuals, contracts, catalogs, policies.
3. Direct text — Paste any content into the text box.

Important: You don't need to wait for training to finish. The agent starts answering with the first indexed pages while crawling continues in the background. This is the opposite of what Chatbase does — they block the agent until 100% complete.

---

Step 4 — Get your embed code (10 seconds)

After adding a source, the embed screen appears immediately.

<script
  src="https://simple-agent.me/widget/loader.js"
  data-agent-id="your-agent-id-here"
  async
></script>

This 5-line snippet is all you need to paste before the </body> tag on your site.

Technical details that matter:

• Loaded via requestIdleCallback — zero LCP impact on your page
• 3KB initial bundle (Shadow DOM + Preact) — 10× smaller than Chatbase widget (45KB)
• Full Shadow DOM isolation — no CSS conflicts with your site

---

Step 5 — Test before going live

The dashboard has a built-in chat playground right in the configuration panel. Before copying the script, ask 3 questions about the content you just trained:

1. An obvious question (answer should be clearly in the source)
2. An edge case (something vague or poorly documented)
3. An out-of-scope question (to see how gracefully the agent declines)

The color dot on the agent avatar shows confidence:

• Green — high-confidence answer, well-anchored in the source
• Amber — partial answer, agent signals uncertainty
• No dot / handoff offer — agent doesn't know and suggests talking to a human

---

What happens under the hood

When you add a URL, the pipeline is:

1. Crawl via Firecrawl (respects robots.txt)
2. Chunking by semantic sliding window (512 tokens, 64 overlap)
3. Embedding multilingual with Cohere embed-v4 1024d
4. Indexing in pgvector HNSW for similarity search
5. BM25 tsvector for keyword search (hybrid)
6. Heuristic FCS for per-response confidence scoring

When answering a user, the system runs Cohere Rerank v3 on the top-20 chunks and selects the top-8 most relevant. This is why response quality is far superior to simple RAG systems.

---

Quick comparison: Simple Agent vs alternatives

Aspect	Simple Agent	Chatbase	Botpress
Setup time	< 60s	15–20min	2–4h
Widget size	3KB	45KB	~60KB
Answers during training	Yes	No	N/A
Inline citations	Yes (chunk-level)	No	No
Pricing	Flat $49/mo	Variable credits	Per node/op
Data residency options	Yes	US-only	Depends

---

Next steps after initial setup

Once your agent is running, three adjustments will maximize quality:

1. Add more sources The more relevant content, the better the agent. Add FAQs, return policy, product manual — anything your support team regularly references.

2. Review unanswered questions The dashboard shows conversations where the agent didn't know the answer. Each one is an opportunity: click + Add answer inline and it automatically becomes a new source.

3. Calibrate voice tone In the agent panel, there are 3 presets: Formal, Balanced, and Casual. For e-commerce and customer support, Balanced usually converts better. For premium or B2B brands, Formal. For young, casual products, Casual.

---

FAQ

Do I need a credit card to enter? No. Early access is invite-based while self-serve billing stays blocked.

Will the agent make up answers not in my source? By default, the FCS score filters low-confidence responses and the agent offers human handoff. You can configure it to respond strictly within scope only.

Can I use my own domain for the agent? Yes, starting from the Agency plan ($399/mo). Cloudflare for SaaS manages SSL automatically.

---

Conclusion

Building a quality AI agent no longer requires weeks of development or an enterprise budget. With the right stack, you go from zero to a live widget in about 90 seconds.

The differentiator isn't setup speed — it's response quality. Chunk-level citations, visible confidence scores, and answers during training are details most tools still don't have.

Join early access →