Hey again,
In Part 1, we got introduced to agents - those smart, tireless digital workers that follow instructions, make decisions, and help automate tasks. We spoke about how they work, what makes up their core (tools, memory, reasoning), and when to actually consider building one for your workflow. If you missed it, no worries - you can catch up on that story here.
Let’s continue from where we left off. This time, we’re diving deeper - how to actually build one.
Picking the Right Brain
Before we let our agent loose in the wild, we need to give it a brain - and not just any brain, but one that fits the job.
Different models are good at different things. Some are fast and cheap but not great at complex tasks. Others are slow, expensive, but can reason like a top consultant. Here’s a simplified table.
Not All Heroes Are Proprietary
If you’re someone who loves control or wants to avoid big-model pricing, open-source models might be your best friend.
These models are like open toolkits - tweakable, transparent, and yours to shape. They don’t come with the bells and whistles of GPT-4, but in the right hands, they’re powerful.
Here are some names you might want to explore later:
- LLaMA 3 (Meta): Versatile and widely used
- Mistral 7B: Lightweight and efficient
- DeepSeek-R1: Huge model with deep reasoning
- BLOOM: Great for multilingual support
- GPT-J: Classic open-source generalist
- Qwen 3, Gemma, Falcon 180B - each with their own flavor
We’ll go deeper into open-source agents in another letter. For now, just know that you don’t have to depend on the big guys.
Let’s Talk Tools
An agent without tools is like a chef without a knife.
Tools allow agents to do things - fetch data, send emails, talk to CRMs, or even control your screen. They’re usually APIs, scripts, or even mini-agents themselves.
There are 3 big buckets tools fall into:
- Data tools – For fetching info. Think: “What’s this user’s payment history?”
- Action tools – For taking action. “Send this email,” or “Update that record.”
- Orchestration tools – For managing other agents. Like a manager overseeing a team.
The key to a good tool? Make it reusable, tested, and predictable.
Setting Clear Instructions
Now imagine our agent’s first day at work.
You’d want to give them a clear, step-by-step instruction manual. That’s what your prompts and routines are. If your instructions are vague, your agent will stumble.
Here’s what helps:
- Use existing docs - Turn SOPs or customer service scripts into prompts.
- Break it down - “Find order ID → Check refund status → Send message”
- Be specific - Include exact actions like “Ask for mobile number,” not just “Check user.”
- Handle weird cases - What if the user doesn’t respond? Or gives an invalid email?
And Lastly… Think Bigger
You can start with one agent. But imagine a small team - one for emails, one for customer support, one for lead qualification. Agents can manage each other, pass tasks, and collaborate like a team in your office.
We’ll talk about multi-agent systems and setting rules for collaboration in Part 3.
But today, take this away: Pick the right model, Build helpful tools, Write crystal-clear instructions.
You’re not just building an agent. You’re training a digital teammate.
Until next time,
Jishnu