Harris is a marketing veteran I met in Singapore. He spent almost five years at Apple's online store, moved into advertising at Nanigans, a Facebook Marketing Partner, where he ran Greater China and managed overseas campaigns for leading strategy game publishers, then joined Meta to cover games in the region. Lilith, miHoYo and Bilibili were among his clients.
"At the peak, nine of the world's top ten strategy game companies were from China."
He left Meta in 2022 to start a company, went through blockchain games, and went all in on AI at the end of 2024 with A65 in Singapore. A65's ambition is large: a cluster of AI agents that takes over a company's whole marketing chain, from customer insight to media strategy to ad creative to round-the-clock monitoring.
Don't build agents for the sake of agents
At startup events in Singapore, he asks the room the same question:
"If you had an intern position today, could you clearly define what that intern would do for you? If you can't, you can't define a good agent either."
A65 made this mistake early. They used agents as productivity tools, for vibe coding and to speed up staff. They soon realised that if agents are just tools, the product itself doesn't change, you're only faster. So they flipped it: the agent is the product.
No memory in functional agents
This design decision stayed with me. A65 has three kinds of agents: AI employees, functional agents that execute specific skills, and governance agents that evaluate and monitor.
Functional agents deliberately have no memory.
"I can't let client A's information affect client B's output. And client B can never see client A's data."
When they first ran multiple clients' briefs through agents with memory, the agents' skills and memory drifted toward whichever client used them most. To serve hundreds or thousands of clients, consistency and stability are the hard part. So A65 spends far more on governance and guardrails than most people would expect of an agent company.
If you want 100 from AI, bring 85 yourself
We talked about CreativeOS, their ad creative system. I mentioned that many marketing friends don't use AI for content because the output feels mediocre.
"It's not that AI can't do it. Your prompt and your input aren't good enough, so the output isn't. If you want AI to get you to 100, you have to be at 85 yourself."
So instead of letting users type "make me a coffee ad", CreativeOS builds a professional creative workflow into the system: find the customer's motivation, combine it with the product's selling points, form a strategy, break it into ideas, then storyboard, hook, call to action and brand message. A creative director with ten years' experience and a beginner get very different results from the same system.
91.6% is enough
A65's CrowdOS builds synthetic audiences: AI agents that simulate consumers for market research. Each agent's personality comes from the Big Five model, and every day the agents are fed news and left to debate and form views. Calibrated against Pew Research's public panels, agreement is 91.6%.
I asked about the remaining 8.4%. He gave three reasons not to chase it. The calibration targets past reports, and matching the past perfectly doesn't mean predicting the future better. Statistically, anything above 75% is already usable. And every fraction of a point beyond 91.6 costs disproportionately more.
"We're not doing perfect scientific research. We're building a commercially usable product."
That restraint is rare in AI, where the usual story is that more accurate is always better.
They don't steer the agents' opinions. "We only feed them raw material," he said. But they watch for any agent so persuasive that it converts all the others. To reduce the bias of any one model toward English-speaking, wealthier, urban, very online people, they run five model families side by side, from US and Chinese labs, and calibrate each market separately. Not every country has Pew-grade data, so some markets are filled in from social media.
A 90-point product for Tier 1 is a 120-point product for Tier 2
Their large clients get custom work at 100 points. Standardise that and you lose the custom 10%, leaving 90.
"But that 90-point standard solution, given to a mid-sized Tier 2 client, might be a 120. They're nowhere near as advanced."
It isn't a new idea, but I hadn't heard it put so clearly for agent products: let your top clients train the product, then release that concentrated know-how to the middle of the market.
Fluency is not capability
A65's website carries the line "fluency is not capability". It comes from ELIZA, the 1965 MIT chatbot, which is also where the 65 in A65 comes from.
He is pragmatic about technology choices. Some things they hard-code. MCP is popular, but where they see its weaknesses, they call APIs directly.
"An API isn't wrong just because it's older. What's the most efficient, lowest-cost way to solve the problem? We do that."
Most AI startups tell a story where newer is better and more agentic is better. Harris tells one about going back to the fundamentals.
*This article draws from a podcast conversation with Harris Wang, founder of A65 (a65.ai).