Last night I attended the Agents & APIs Developer Meetup hosted by Postman at Capital Factory in Austin, TX.
Four speakers. Four completely different angles on AI agents. One unforgettable evening.
Before I get into the technical details, I want to acknowledge Josh Bear, the CEO and cofounder of Capital Factory, who we recently lost in a tragic plane crash. The meetup was held in his honor. 🙏
Speaker 1 — Talia Cohen (Postman)
"In 2026, Engineers Don't Write Code. They Understand Impact."
Every AI tool today can write code. That's solved. The hard part is understanding what breaks when you change something.
Postman launched their AI Engineer powered by a Context Graph — a continuously updated map of your entire API ecosystem.
The Demo:
Talia renamed one field — ID to employee_ID. The AI came back with 13 artifacts that needed updating across two repos. One tiny change. 13 things broken. Including a payroll service in a completely separate repo she wouldn't have thought of.
Key Quote:
"In 2026, engineers don't need to write code. They need to understand the impact."
Speaker 2 — Jacob Rothfuss (LibLab/Postman)
"AI Agents Don't Need More Context. They Need Better Context."
There is a research backed concept called the Dumb Zone. Past 50% of maximum context window — AI performance drops sharply.
When you give an AI agent raw API docs — it pulls 10,000 lines of stuff it doesn't need. Give it an SDK instead.
The Experiment:
Same app built twice — with and without SDK.
| Without SDK | With LibLab SDK | |
|---|---|---|
| Iterations | 4-5 | 1 |
| Quality | Junior level | Senior level |
| Time | Long | 5 minutes |
| Did it work? | No | Yes |
LibLab generates SDKs from any OpenAPI spec in 5 seconds.
Key Quote:
"The barrier to entry for software has never been lower. But give AI an SDK — not raw API docs."
Speaker 3 — Dave R (Coder)
"AI is Moving at Machine Speed. Your Governance Isn't."
Since January 2026 — 35 to 40 AI security incidents. That is 6-7 per month. Data exfiltration. Data deletion. Cost overruns.
The Framework — See it → Control it → Scale it:
Migrate — Get agents off laptops into managed environments so you can see them
Modernize — Add rules, cost limits and access controls around them
Multiply — Once controlled, scale up with specialized agents confidently
5 Must Haves for Your AI Execution Layer:
| Must Have | Purpose |
|---|---|
| Centralized orchestration | Know what every agent is doing |
| Workspace isolation | Each team gets only what they need |
| Policy as code | Consistent rules via Terraform |
| LLM gateway | Record every agent to model call |
| Human in the loop | Verify before agents act alone |
Key Quote:
"Go small. Eat the elephant one bite at a time. Make sure there is a human in the loop."
Speaker 4 — Clement Benschev (PostHog)
"Why Not Just Let Things Build Themselves?"
Contrarian view — go fast, automate everything, observe everything.
His Demo:
Built a Telegram betting bot that:
- Tells you World Cup matches
- Gets live betting odds from Polymarket
- Creates betting pools with QR codes
- Creates its own API endpoints dynamically
- Every new endpoint automatically gets a feature flag as guardrail
PostHog Core Features:
| Feature | Purpose |
|---|---|
| Session Replay | Watch exactly what users see |
| Feature Flags | Safe on/off switches for rollouts |
| Product Analytics | Understand user behavior |
| Scouting Agent | AI finds bugs and prepares fixes |
| PostHog Code | AI product editor with code and user data access |
| AI Observability | Every agent action tied to user ID and session ID |
Key Quote:
"Take risks and have fun building whatever is next."
How All Four Talks Connect
| Speaker | One Line Summary |
|---|---|
| Talia — Postman | Give AI context of your entire system |
| Jacob — LibLab | Give AI clean SDKs not raw APIs |
| Dave — Coder | Control and govern your AI agents |
| Clement — PostHog | Observe and measure everything |
My Top 5 Takeaways
- Context beats code — right context matters more than code generation
- Always use SDKs — cheaper tokens, better output, fewer iterations
- Governance is urgent — 35-40 incidents since January is not a drill
- Feature flags are your best guardrail for autonomous agents
- Human in the loop — build trust gradually before going fully autonomous
Final Thought
We are living through the most exciting period in software engineering history.
But excitement without infrastructure is just chaos.
The developers who win in this era won't just build the coolest agents. They will build the right infrastructure around those agents.
That is the real lesson from tonight.
Huge thanks to Patricia Dugan and the Postman team for organizing. RIP Josh Bear — we honored you tonight. 🙏
Follow me for more content on AI, Java, Spring Boot, and enterprise architecture.