build an AI agent with me from scratch
30-minute guide to building an AI agent that works.
👋 Hey, I’m Suhas and welcome to this week’s edition of the newsletter. Each week I tackle questions about building products, driving growth, and accelerating your career.
I’ve been building AI agents for the past few weeks, and it’s been one of the most interesting experiments I’ve run.
Most people think agents are just “ChatGPT with extra steps.” That’s not true.
Today I’ll show you what makes agents different and how to build one.
What is an AI agent?
An AI agent is a system that uses an LLM (like Claude or ChatGPT) to make decisions and take actions based on rules you define. repeatedly and autonomously.
Think of it like this:
ChatGPT waits for you to prompt it, then responds.
An agent follows a set of instructions you've given it, applies those rules to new inputs, and takes action, without you being involved each time.
What makes agents different:
Persistent memory: Your rules and criteria stay loaded permanently. Open it a month later, it still knows exactly how you want things done.
Tool use: Agents can connect to external systems—your CRM, databases, APIs. They don’t just suggest “check your database”—they actually check it.
Multi-step execution: One input triggers a chain of tasks. “Read lead → Check database → Score → Send to Slack → Update spreadsheet” happens automatically.
Trigger-based: Agents can watch for events. “Every time a new lead comes in, evaluate it.” You’re not involved.
This is why I’m obsessed with agents right now. They’re smart, operational and have high agency.
What we’re building
A lead qualification agent for B2B SaaS.
I wanted something that could:
Read inbound leads
Score them 1-12 based on specific criteria
Categorise them (Hot Lead / Nurture / Disqualify)
Explain reasoning
Tell me exactly what to do next
Platform: Claude Projects (free, you already have access)
Why Claude Projects over other tools?
I tested zapier, n8n, and custom GPTs. Claude Projects won because:
No setup complexity
Persistent custom instructions
Knowledge base support
Step 1: Create your agent's workspace
Open Claude → Projects → Create Project
Name it “Lead Qualification Agent”
What just happened: You created a persistent workspace. Everything you configure here stays. Open it 3 months later, all your rules are still there.
Step 2: Write your agent’s brain
Click “Set custom instructions”
This is where you codify your decision framework. Here’s mine:
Why this structure works:
Point-based scoring = Auditability When someone asks “Why did we pass on this?”, you have data, not vibes.
Forced reasoning = No black boxes The agent must show its work. You can audit every decision, just like code reviews.
“What would improve this” = Feedback loop Mid-quality leads (5-7 scores) are where most time gets wasted. This section tells you and the lead exactly what needs to change.
Click “Save”
Step 3: Test it
Open a new chat in your project.
Paste a real lead. Here’s one I tested:
Hit enter.
The agent’s response:
This is where it clicked for me.
The agent showed its work, gave me a ready-to-send response, and told me exactly what would turn this into a hot lead.
Where it broke for me
I tested 20 more leads over the past week.
18 were perfect, but it got 2 of them completely wrong.
One failure that was good feedback:
A 200-person e-commerce company inquired about the enterprise plan.
Budget: $10K/month.
Agent score: 6/12 - “NURTURE”
I was confused. $10K/month is a dream lead, why nurture?
Then I looked at my prompt:
Established company (100+ employees OR known brand): +3 points
The agent followed my rules perfectly, but the rules were incomplete.
I’d been unconsciously making an exception: “If budget is $5K+ monthly, company size is less important.” but I never wrote that down.
This is a key insight into building agents: They expose gaps in your own thinking.
If your decision framework has holes, the agent falls through them. Every edge case reveals an implicit rule you’ve been applying unconsciously.
So I updated the rule and re-ran the lead.
Established company (100+ employees OR known brand OR budget $5K+ monthly): +3 points
New score: 9/12 - "HOT LEAD"
What else can this agent do?
Once you have the basic scoring working, you can extend it:
Connect it to your CRM Upload a CSV of past leads (company names, deal sizes, outcomes). The agent can reference this when scoring new leads.
Add competitive intelligence Upload competitor comparison docs. When a lead mentions a competitor, the agent can suggest specific differentiation points.
Multi-language support Leads come in Hindi? The agent scores them using the same criteria, responds in Hindi.
Integration with Slack/Email Set up forwarding. Every new lead gets auto-scored, posted to your #leads channel with the agent’s analysis.
What I'm building next
I’ve been running this agent for 2 weeks and the I’ve been saving about 2.5 hours per week just on lead scoring.
And the bigger unlock is consistency. Every lead gets evaluated the same way, no matter who’s tired or what mood they’re in.
Next steps:
I'm making this production-ready and testing it with a few SaaS founders I know.
IN AI THIS WEEK
Emergent raises $70M Series B led by Khosla Ventures. Read more
The AI vibe coding platform hit $50M ARR in just 7 months with 5M+ users building full-stack apps using natural language. One of the fastest Series A to B timelines in AI history.Remotion Skills shifts video creation from code to instructions. Read more
The React-based video framework now lets AI agents (like Claude Code) generate professional animations from natural language prompts.Apple is turning Siri into an AI chatbot powered by Gemini. Read more
The new Siri will replace today’s interface and live deeply inside iOS 27, iPadOS 27 and macOS 27, letting users chat, search the web, generate content and images, and act on on‑screen/contextual data across core apps.
Upcoming Events
Learn How Voice AI Products Are Built
Hyderabad | Jan 24
Register Here
From 0 → IPO: Real Product Lessons from a PW PM
Delhi | Jan 31
Register Here
The Product Folks’ Bits n Atoms
Chennai | Jan 31
Register Here
Jobs Jobs Jobs
Senior Product Manager
Lead Product Manager
Director Product
Roles open across top companies like PhonePe, Swiggy, Groww & many more.
If you've already built an agent, drop a comment.
I'm curious what use cases are working for you.
Have a great week,
Suhas 👋
Are we categorising Claude Projects as agents solely because of Skills and MCPs?