AI Agents in Production: What 650 Builders Taught Me About Cost, Memory, and Trust

🚀 AI agents are no longer on slides. They're in production - running real work, making real decisions, and moving real money. Last week I saw it firsthand when 650 people from 1,300 signups filled the space for AI Agent Night, Hermes Edition, our third event of this kind, put together by Lionel Sim and The AI Capitol community. (And yes, we hosted it at the AWS Singapore office.)

Seven speakers, seven real systems

Completely different backgrounds. All shipping real things.

• 🔹 Bryan Chua (Gopomelo) - running a multi-tenant fleet of agents in production, with token caps and security locked down from day one.
• 🔹 Nathaniel Ng - spec-driven development with Kiro: get your requirements right and the code follows.
• 🔹 Benjamin Cheng (Sandpiper) - a finance director, not an engineer, automating debtor tracking and compliance questionnaires… on a 10-year-old laptop with a broken screen.
• 🔹 Marcus Cheu (Notion) - turning a personal agent into team capability by giving everyone one shared 'second brain.'
• 🔹 Amarnath R. (Airwallex) - agents that touch real money safely, with server-side human approval before anything moves.
• 🔹 Jasper Hartono - sent a WhatsApp from a pair of AR glasses just by looking and speaking. The room's favourite by a mile. 🕶️

The question has changed

💡 Here's what struck me: nobody was asking 'can agents do this?'

They were asking 'how do I run this responsibly - without my token bill exploding, without losing the thread on long tasks, and without handing over keys I can't take back?'

Cost. Memory. Trust.

The same three worries surfaced in every single talk. That's a technology growing up.

What it means to ship agents responsibly

⚖️ When Benjamin - a finance director with no engineering background - can automate compliance workflows on a decade-old laptop, we're past the 'proof of concept' phase.

When Amarnath's agents touch real money but require human approval before anything moves, we're designing for accountability, not just automation.

When Bryan's multi-tenant fleet has token caps and security baked in from day one, we're building systems that scale without losing control.

This is what production looks like. Not perfect. Not hype. Just real teams solving real problems with guardrails they trust.

The human-first throughline

🎯 Every speaker in that room understood something fundamental: agents amplify human judgment. They don't replace it.

Marcus didn't build an agent to eliminate his team's thinking - he built one to give everyone a shared second brain. Amarnath didn't automate money movement blindly - he put a human checkpoint exactly where it matters most.

That's the shift. We're not asking agents to be autonomous in the sci-fi sense. We're asking them to be reliably useful - to handle the repetitive, the tedious, the high-volume - so humans can focus on the judgment calls that actually need us.

What are you building?

If you're working on agents right now - in production or getting close - I'd love to hear what you're learning. What's your biggest worry? Cost? Memory? Trust? Something else entirely?

The builders in that room weren't waiting for permission. They were shipping, learning, and sharing. That's the energy that moves this work forward. 🙏

More soon.

#AlwaysDay1 #AIAgents #AgenticAI #ProductionAI #AILeadership #FutureOfWork