The tool landscape shifted under our feet
The loudest grumble on the floor was not about models — it was about Claude Code rate limits and pricing. Enough builders have been pushed toward Cursor or OpenAI Codex that the competitive map is redrawn weekly. Several people said they only opened Cursor because of Claude friction — then stayed because recent 3.x updates, shaped by Ryo Lu's design team, feel stable and genuinely pleasant to work in.
Anthropic felt absent in the social layer. I did not see Claude sessions dominating hallway chatter the way Codex and Cursor did. Gabriel Chua, leading the Codex community in Singapore, always had a crowd waiting to talk. Whether Anthropic is choosing research-house seriousness over community theatre is a fair question — Rajesh Acharya's take on LinkedIn matched what I felt: they may simply not treat dev-tool evangelism as core mission.
On stage, Thibault Sottiaux's Codex talk and Jimmy Lai's Next.js session anchored the "ship what's next" narrative. In the room, the vibe was execution: fewer manifestos, more working agents.
Side events were the main event
Credit where it is due: fellow Cursor ambassadors Agrim Singh, Brian C., and Sherry Jiang ran an organisational machine. Frequent emails, a thick side-event roster on aievents.sg, and multiple high-calibre meetups any single evening — some as good as the main-track talks.
Highlights from the orbit:
- Build with Cursor Singapore — fireside chat plus hands-on building with the Cursor team.
- Daytona AI Builders w/ Zo Computer — solos and startups demoing real agent workflows; Geoffrey Huntley's warning on how unit economics of software businesses have changed landed hard.
- Personal Agents Demo Night and Zo Computer sessions — the closest community-demo quality I have felt since HCMC AI Tinkerers (shoutout to Yong Quan Tan, Son Le, and Edward Tran).
- ClawCon Singapore at AWS — still packed, festival energy, despite OpenClaw's reputation taking hits in builder circles.
- Hermes Night — the personal-agent conversation has a new default name in many rooms.
Production agents: what I am taking back to work
Vedran Jukic's Daytona talk — Why Sandboxes Are Non-Negotiable for Autonomous AI Agents — matched what I see in client work. Daytona looks like the best escape valve for production agents outside hyperscaler walled gardens. Every attendee at the builders event received Sam Bhagwat's agent book; if you missed a copy, grab Principles of Building AI Agents from Mastra.
My current stack hypothesis from the week: Daytona plus the Mastra framework for teams that need agents in prod with evals, traces, and sane defaults — Mastra's workshop title literally walked from "first agent" to "production claw." For non-technical colleagues who will never touch deployment or security, Zo Computer kept coming up as the most approachable way to put an agent in someone's hands without becoming their unpaid SRE.
At Cipher Projects, the company I founded, that split maps cleanly: sandboxed execution and orchestration for engineering teams; guarded, product-shaped surfaces for everyone else. Bear Consulting Group sits adjacent in the same ecosystem — helping cloud and AI teams operationalise the same patterns inside enterprises that do not want to learn fifteen new vendors in one quarter.
OpenClaw, Hermes, and the personal-agent wars
OpenClaw / Claw culture still has a hardcore fanbase — ClawCon filled a room — but hallway sentiment skewed negative on stability: endless troubleshooting, breaking changes, operator fatigue. Meanwhile Hermes has become the name people reach for when they mean "personal agent that might stay up this week." The fragmentation is exhausting; the winner will be whoever nails reliability and onboarding, not manifestos.
Designers owned the room
The most electric main-stage moments were often design talks, not infra keynotes. AI has pulled designers (and design-minded engineers) back toward code as a creative outlet — and it shows in how they present.
Ryo Lu's Designing the Next Cursor, Josh Newton from Microsoft, Sabina Cabrera from Magic Path, Phil Hedayatnia on Airfoil, Annie Luo on friction worth keeping — these sessions got people leaning forward. If you work in platform or ops, ignore design track at your peril: the UX of agent products is the adoption curve.
Worth your calendar: Melbourne in June
If you work in AI engineering and your employer still funds learning, get tickets for AI Engineer Melbourne in June. Singapore set a high bar; the same 65labs crew knows how to stack workshops, speakers, and side events. Especially worth it if someone else is paying.
Quick reference — official program shape
For readers planning their own trip:
- Friday 15 May — Workshops + Leadership track at SMU (Cursor, Mastra, OpenAI Codex, Vercel, Convex, LlamaIndex, and others).
- Saturday–Sunday 16–17 May — Talks at Capitol Kempinski across Software, Design, and Physical AI tracks.
- Producers — 65labs and Software 3.0 Inc; community coordination heavily driven by ambassadors and sponsors.
Singapore's Minister for Foreign Affairs opened with a "second brain" framing for national AI adoption — a reminder that these conferences are not only about npm packages. They are about where governments, enterprises, and solo builders align on risk, sovereignty, and speed.