> ## Documentation Index
> Fetch the complete documentation index at: https://aiplaybooklac.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Build

# Building AI Agents on an LAC Small-Business Budget

> **Created by Adrian Dunkley** | [maestrosai.com](https://maestrosai.com) | [ceo@maestrosai.com](mailto:ceo@maestrosai.com) | Fair Use

***

You do not need a developer to ship a useful agent in 2026. You do need to pick the right build path. This page walks the three practical paths from cheapest and simplest to most flexible, with real USD pricing that works for LAC micro and small businesses.

Read [design.md](design.md) first so you know what you're building.

***

## Pick your path

| Path         | You need                            | You get                                           | Monthly cost | Build time    |
| ------------ | ----------------------------------- | ------------------------------------------------- | ------------ | ------------- |
| **No-code**  | A browser, a credit card            | A working agent inside Claude, ChatGPT, or Gemini | $0 to $20    | 1 to 4 hours  |
| **Low-code** | Willingness to wire blocks together | An agent across WhatsApp, email, Sheets, CRM      | $20 to $60   | 4 to 16 hours |
| **Code**     | Python or JavaScript                | A custom agent you own and can host anywhere      | $20 to $200  | 8 to 40 hours |

Start at the simplest path that covers your use case. You can always graduate.

***

## No-code path

This path runs entirely inside a consumer AI product. No servers, no SDKs, no code.

### Options

| Platform                                 | What it is                                                 | Best for                                    | Cost                        |
| ---------------------------------------- | ---------------------------------------------------------- | ------------------------------------------- | --------------------------- |
| **Claude Projects and Skills**           | Persistent context + custom skills in the Claude web app   | Bilingual writing, research, business docs  | Claude Pro \$20/mo          |
| **ChatGPT Custom GPTs**                  | Custom GPT with instructions, knowledge files, and actions | Customer FAQ, sales scripts, language tutor | ChatGPT Plus \$20/mo        |
| **ChatGPT workspace agents**             | Team-shared agents that run across ChatGPT and Slack       | Small team back-office flows                | Business plan, \$25/user/mo |
| **Gemini Gems**                          | Custom Gemini personas with Google Workspace integration   | Google Docs, Sheets, Gmail flows            | Google AI Plus \$20/mo      |
| **Microsoft Copilot Studio (light use)** | Low-code flow builder inside Copilot                       | Office 365 users                            | M365 Copilot \$30/user/mo   |

### What you can and can't do

**Can**: Upload knowledge documents, set a persona, add custom instructions, call web search, run basic code. Share with teammates on paid tiers.

**Can't**: Reliably call your own WhatsApp, run on a schedule, or plug into your own CRM without a paid integration tier.

### A 60-minute no-code build

Goal: a Spanish-speaking customer-service agent for a Mexican bakery.

1. Open ChatGPT, click "Create a GPT".
2. Paste a goal and style guide in Spanish (use the template from [design.md](design.md)).
3. Upload your menu, hours, FAQs, and delivery zones as knowledge files.
4. Add 5 example conversations under "Conversation Starters".
5. Test with 10 real customer messages. Fix the instruction on any miss.
6. Share the link with your staff. Done.

Recurring cost: \$20/month. Time to ship: one hour.

***

## Low-code path

This path lives outside the AI product and glues the agent to the tools your business already uses.

### Options

| Platform                       | Strongest at                                 | Free tier              | Paid                   |
| ------------------------------ | -------------------------------------------- | ---------------------- | ---------------------- |
| **n8n**                        | Self-hosted automations with branching logic | Yes, self-host is free | Cloud \$20/mo          |
| **Make (formerly Integromat)** | Visual scenarios with many integrations      | 1000 ops/month         | Core \$9/mo            |
| **Zapier with AI steps**       | Fastest integration count                    | 100 tasks/month        | \$20/mo starter        |
| **Relay.app**                  | Agent-first design with human-in-the-loop    | Free tier              | \$9/mo                 |
| **Pipedream**                  | Developer-friendly, JS/Python steps          | Yes, 10k invocations   | \$29/mo                |
| **Microsoft Copilot Studio**   | Flows inside Microsoft 365                   | Trial                  | \$200/month per tenant |

### The typical low-code agent

A low-code agent is usually a loop:

1. **Trigger**: WhatsApp message received, email arrived, row added to a sheet.
2. **LLM call**: Pass the message to Claude or GPT with a system prompt.
3. **Decide**: Parse the LLM output. Branch on the intent.
4. **Tool step**: Call WhatsApp Business API, Sheets, HubSpot, a payment API.
5. **Reply or log**.
6. **Loop or stop**.

### Sample pricing for an LAC SMB

| Component                      | Example                           | Monthly USD    |
| ------------------------------ | --------------------------------- | -------------- |
| Workflow platform              | n8n Cloud starter                 | \$20           |
| WhatsApp Business API          | 360dialog or Twilio, low volume   | \$15           |
| LLM API                        | Claude Haiku 4.5 at \~1000 convos | $3 to $10      |
| Google Sheets                  | Free                              | \$0            |
| HubSpot Starter CRM (optional) | Starter plan                      | \$20           |
| **Total**                      |                                   | **$38 to $65** |

That is within reach for a tour operator, salon, clinic, or small exporter.

### A low-code starter: WhatsApp booking agent in n8n

1. Connect WhatsApp Business API to n8n as a trigger.
2. Pass the incoming message to a Claude Haiku 4.5 node with a booking-specific system prompt.
3. If the LLM returns intent "BOOK", call a Google Calendar node to check availability.
4. If free, call a Stripe (or PayPal) node to generate a payment link.
5. Send the link back via WhatsApp.
6. On payment webhook, add the booking to the calendar and send confirmation.
7. If the LLM returns intent "OTHER", forward the message to the owner's phone.

You can have this running in a weekend.

***

## Code path

This path gives you full control: custom tools, self-hosting, private deployment, offline-capable agents, and compliance with LGPD, LFPDPPP, and similar LAC data rules.

### SDKs and frameworks in 2026

| SDK / framework              | Language           | Strongest at                                        | License        |
| ---------------------------- | ------------------ | --------------------------------------------------- | -------------- |
| **Claude Agent SDK**         | Python, TypeScript | Tool-using agents with skills, memory, computer use | Commercial API |
| **OpenAI Agents SDK**        | Python, TypeScript | Multi-agent workflows, handoffs, Responses API      | Commercial API |
| **Gemini Enterprise agents** | Python, TypeScript | Google Cloud + Workspace integrations               | Commercial     |
| **LangGraph**                | Python, TypeScript | State-machine-style agents, graph execution         | MIT            |
| **LlamaIndex agents**        | Python             | Retrieval-heavy agents                              | MIT            |
| **CrewAI**                   | Python             | Role-based multi-agent teams                        | MIT            |
| **Microsoft AutoGen**        | Python, .NET       | Research multi-agent patterns                       | MIT            |
| **Semantic Kernel**          | Python, .NET, Java | Enterprise .NET stacks                              | MIT            |

### When to use which

* **Claude Agent SDK**: if the agent needs long-running tasks, computer use, or skills. Default choice for most business agents.
* **OpenAI Agents SDK**: if you're already on OpenAI, want their Responses API, or need multi-agent handoffs.
* **LangGraph**: if you want provider-agnostic code and explicit control over state and branching.
* **CrewAI**: if you want to prototype a multi-role team (researcher + writer + reviewer) in an afternoon.

### Hosting options for LAC

| Host                                                | Good for                                | Pricing note        |
| --------------------------------------------------- | --------------------------------------- | ------------------- |
| **Render, Fly.io, Railway**                         | Small agents, simple deploy             | $5 to $25/mo        |
| **Vercel, Cloudflare Workers**                      | Event-driven agents                     | Generous free tiers |
| **AWS São Paulo (sa-east-1), Santiago (sa-west-1)** | LAC data residency                      | Pay as you go       |
| **GCP São Paulo, Santiago**                         | LAC data residency                      | Pay as you go       |
| **Self-hosted on a small VPS**                      | Full control, lowest cost               | $5 to $20/mo        |
| **Local Mac Mini or laptop with Ollama**            | Privacy-first, offline, SLM-based agent | One-time hardware   |

For data-sensitive agents under LGPD or Jamaica DPA, prefer São Paulo or Santiago regions. See [governance/README.md](../governance/README.md).

### A minimal Claude Agent SDK example

```python theme={null}
from anthropic import Anthropic
from anthropic.lib.agents import Agent, tool

client = Anthropic()

@tool
def check_calendar(date: str) -> str:
    """Return available slots for a given date."""
    return "10:00, 14:00, 16:00"

@tool
def send_whatsapp(to: str, message: str) -> str:
    """Send a WhatsApp message via the Business API."""
    return "sent"

agent = Agent(
    model="claude-sonnet-4-6",
    system="You book tour reservations for a Jamaican tour operator. "
           "Reply in the customer's language. Always confirm before booking.",
    tools=[check_calendar, send_whatsapp],
    max_steps=15,
)

response = agent.run("Hi, do you have spots for 2 on Saturday at noon?")
print(response.final_message)
```

That's under 30 lines and it's a real agent. Swap the tools for your own and you have a working booking flow.

***

## Cost math the LAC business owner should do

Before you build, do this math:

1. **Hours saved per week** = H
2. **Your hourly value in USD** = V
3. **Monthly platform + API cost** = C
4. **Monthly savings** = H \* 4 \* V
5. **Keep only if**: monthly savings > 3 \* C

Example. A Panamanian accountant spends 8 hours/week on invoice entry. Her hourly value is $25. An invoice-extraction agent costs $45/mo to run.

* Savings: 8 \* 4 \* 25 = \$800/mo.
* Cost: \$45/mo.
* Ratio: 17x. Build it.

A corner-shop owner spends 1 hour/week on WhatsApp replies. Her hourly value is $8. A WhatsApp agent costs $40/mo.

* Savings: 1 \* 4 \* 8 = \$32/mo.
* Cost: \$40/mo.
* Ratio: 0.8x. Skip it. Use a free Custom GPT instead.

***

## Security, privacy, and compliance basics

Before you ship to customers:

* **Never let the agent execute actions without a confirmation step** for anything that sends money, deletes data, or contacts customers.
* **Log every tool call** for 30 days minimum. This is your audit trail if a customer disputes a booking or a charge.
* **Mask PII** in logs sent to third-party platforms.
* **Comply** with local data-protection law. See [governance/README.md](../governance/README.md).
* **Document the agent**: what it does, what it can't, and who to contact when it goes wrong.

***

## Graduating from one path to the next

You'll know it's time to graduate when:

* **No-code to low-code**: you hit "I wish this could talk to my WhatsApp" or "I need this to run at 2am".
* **Low-code to code**: you hit "I need a custom tool the platform doesn't offer" or "data can't leave my country".

You don't have to rebuild. Most businesses end with a hybrid: a Custom GPT for one use case, an n8n workflow for another, and one Python agent for the thing nothing else can do.

***

## Next step

Pick a pattern from [patterns.md](patterns.md) and build it this week.

***

*Created by Adrian Dunkley | MaestrosAI | maestrosai.com | [ceo@maestrosai.com](mailto:ceo@maestrosai.com)*
*Fair Use, Educational Resource | April 2026*
*SEO: build AI agent | construir agente IA | Claude Agent SDK | OpenAI Agents SDK | n8n agent | agent builder LAC | AI agent WhatsApp Caribbean*