14 Principles of Building AI Agents (Learned the Hard Way)

What I learned by building 50+ AI agents and copying the Multi-Agent Research System by Anthropic. Best practices and mistakes to avoid.

Hey, welcome to the freemium edition of The Product Compass newsletter.

Every week, I share actionable tips, resources, and insights for PMs.

Here’s what you might have missed recently:

1. The Ultimate AI PM Learning Roadmap

2. The Ultimate Guide to AI Agents for PMs

3. Introduction to AI Product Management

4. Base44: A Brutally Simple Alternative to Lovable

5. How to Figma → Jira epics and stories in 10 min.

Consider subscribing or upgrading your account for the full experience:

---

In the recent months, I’ve built 50+ AI agents, experimented with 7+ agentic frameworks, and copied the Multi-Agent Research System by Anthropic.

Here's what I learned (best practices and mistakes to avoid):

---

1. Don’t Use Agents If You Don't Have To

Nobody cares if it's an AI agent or a simple script, as long as it works. A good old if/else is faster, cheaper, and more reliable. And it's often all you need.

Save the agents for when you really need them. They might easily become a liability.

---

2. Small, Specialized, and Decoupled

Think "team of specialists," not "one agent to rule them all." A planner plans. A summarizer summarizes. A verifier checks. Decoupled agents are cheaper to run, easier to test and fix, and way more predictable.

---

3. Enforce Structured Output

I've learned that text is a mess to deal with. JSON is easier to debug, cheaper to parse, and acts like a contract between agents.

Bonus: you can validate it automatically and stop errors before they spread.

---

4. Explain the Why, Not Just the What

I've discovered that anthropomorphizing AI works in many contexts. Here, lead with context not control.

When delegating a task, don't just define the objective. Explain why it matters and provide the context in which you need it.

This helps AI agents make better decisions with shorter prompts.

---

5. Orchestration > Autonomy

Autonomy sounds great, but what you need more in real life is predictability. Move all known logic (if/then, loops, retries, known procedures) out of agent prompts and into the orchestration layer.

---

6. Prompt Engineering > Fine Tuning

Before you jump to fine-tuning, ask: Why is the model failing?

• If it’s missing facts → try RAG.

• If it’s wrong formatting or doesn't follow your brand style → maybe fine-tune. But 80% of the time, it’s just a prompt problem.

Source: AI Product Management Certification by Miqdad Jaffer (OpenAI)

---

7. Double Down on Tool Descriptions

Treat tool descriptions as micro-prompts that guide agents’ reasoning. Unfortunately, descriptions provided by MCP servers are often insufficient and do not consider your specific domain context.

Tell the agent when and why to use the tool, what to avoid, and include examples.

Practical tip: Explain how tools can work in combination (e.g., for the Trello MCP my agent had to list boards, get available lists inside a selected board, etc.) I usually place those instructions in agent prompts.

---

8. Cache Like You Mean It

Often, an agent runs the same task on the same data over and over, like when scrapping a website. Cache responses (e.g., hash of agent ID + input) to reduce latency and API costs.

---

9. Use Shared Artefacts

Do you send documents you collaborate on as attachments? Of course, not. Similarly, empower your agents to collaborate by co-editing shared docs, plans, or code.

Also, often, the next agent doesn’t even need the content of the artefact.

---

10. Log Everything (Seriously)

No logs = no learning. Track everything: inputs, outputs, retries, tool calls, agent thoughts. Add your own app-specific dimensions (e.g., customer type, use case).

Then analyze errors and design evaluators.