Episode 9: Building Your First Agent in CrabShack

If you’ve made it this far in the series, you now understand the basics of AI, automation, agentic AI, and how models, context, and tools fit together. The next step is where things become real: building your first agent. This is where a small business moves from learning about AI to actually using it in a practical workflow.[1]

Start With One Use Case

The first step is to decide what you want to automate. A good first agent should handle one clear business task, not try to do everything at once. Think of repetitive work that is time-consuming, rules-based, and easy to define, such as answering common questions, routing leads, summarizing requests, or pulling data from a system.[2][3] We also covered this in this blog in episodes Five and Seven.

A useful way to choose is to ask three questions:

·         What task happens often?

·         What task follows a repeatable pattern?

·         What task would save time if it was handled faster or more consistently?

If you cannot describe the job in one sentence, it is probably too broad for your first agent.

Map The Tools

Once the use case is clear, list the tools the agent needs to talk to. This may include your CRM, ticketing system, knowledge base, spreadsheet, chat platform, or internal database. The more specific you are here, the easier it is to design the agent without overcomplicating it.[3][4]

This is also where MCP connections come in. MCP helps the agent connect to approved tools in a structured way, instead of relying on fragile one-off integrations. For SMBs, that matters because you want the agent to be useful without turning it into an integration project.[2][3] We covered this topic in more depth in Episode 6.

Keep Humans In Loop

Every first agent should have human checkpoints. The rule is simple: let the agent handle the boring or repeatable parts, but keep people involved when there is risk, ambiguity, or customer impact. That might mean approving outbound messages, reviewing recommendations before they are sent, or escalating anything that looks unusual.[5][3]

A good agent does not replace judgment. It supports it. This is especially important for small businesses, where one bad automation can create more work than it saves.

Choose Model

Next, choose the model that fits the job. Not every agent needs the biggest or most expensive model. A customer support triage agent may need speed and consistency, while a research or drafting agent may need stronger reasoning.[6][2]

The simplest way to think about it is this:

·         Light tasks: use a faster, lower-cost model.

·         Complex tasks: use a stronger model.

·         Sensitive tasks: combine a capable model with strong guardrails and human review.

The model should match the business task, not the other way around. Important considerations are context size  and token pricing. We covered these in our Episode 2.

Write Role File

Now define the agent properly in an Agent-Role.md file. This is where you describe what the agent is for, what it is allowed to do, what tools it can use, how it should respond, and when it must hand off to a human. In practice, this becomes the agent’s operating guide and keeps it aligned to the business job you actually want done.[4][7]

A simple Agent-Role.md usually includes:[7]

·         Purpose.

·         Scope.

·         Allowed tools.

·         Disallowed actions.

·         Escalation rules.

·         Tone and response style.

·         Human approval requirements.

If the use case is the idea, the role file is the rulebook. A good tip here is to use your AI tool to write one for you. Tell it what the agent is called and what its job will be (ie Use XERO MCP to create draft invoices). The AI tool will then write out a detailed Agent-Role.Md file descriptor. Copy this and then paste into the Agent_Role.md File. In CrabShack, this is located in your agents expanded menu and then clicking on Explorer:

Pick The Channel

The final step is choosing where the agent lives. The main options are Discord, Teams, Slack, and Telegram, and the right choice depends on how your team already communicates. If your business already runs in Slack or Teams, keep the agent there. If you want a lighter community-style experience, Discord or Telegram may make more sense.[8][9][2]

The channel is not just a delivery mechanism. It shapes how people use the agent, how often they interact with it, and how naturally it fits into daily work. The best channel is usually the one your team already opens without thinking.

A Simple First Build

A strong first agent for a small business might look like this:

·         Use case: qualify inbound leads.

·         Tools: CRM, calendar, knowledge base.

·         MCP connections: CRM read/write, scheduling, document lookup.

·         Human loop: approve any booked meeting or pricing exception.

·         Model: fast model for triage, stronger model for edge cases.

·         Agent-Role.md: define tone, rules, and escalation.[7]

·         Channel: Slack or Teams for internal use.[8][9]

That gives you a useful, controlled starting point without trying to build a full AI employee on day one.[3][2]