LLMs are central to Manus. Not so for this new AI agent

While big names like Manus are soaking up the AI agent hype, Pokee AI is rethinking how these agents should really work.

The company was founded in October 2024 by Zheqing (Bill) Zhu. A Chinese American with a Stanford Ph.D., he used to lead applied reinforcement learning at Meta.

Unlike its competitors, Pokee isn’t trying to create bigger language models.

Instead, it’s leaning into what Zhu knows best: reinforcement learning (RL), a machine learning paradigm where models learn through relentless trial and error. He believes this approach is the key to smarter planning and decision-making in AI agents.

Like Manus, Pokee helps users automate workflows. It performs tasks such as scheduling meetings and posting social media content across a wide range of web tools. What sets it apart is its ability to chain more tools together in one go. 

“RL can be a very valuable model by itself when you train it properly,” Zhu, who also serves as the company’s CEO, tells Tech in Asia in an interview. “You’re no longer just generating tokens. You have to plan at a very abstract level.”

Essentially, Pokee’s model breaks down a large goal – such as running a marketing campaign – into smaller steps, selecting the appropriate tools to accomplish each one.

For now, investors aren’t waiting around. The startup has raised a US$12 million seed round led by US-based Point72 Ventures, with backing from Qualcomm Ventures, Samsung Next, and heavyweights like Intel board member Lip-Bu Tan and former Adobe CTO Abhay Parasnis.

Pokee is currently in public beta and gearing up for a full public launch in the next month or two. The question is whether a small, RL-powered agent can outmaneuver the giants building with billion-parameter models.

Swapping LLM bloat for RL brains

Zhu’s track record gives his vision some serious credibility. At Meta, his team used RL to optimize recommender systems and ad delivery, which he claims “generated hundreds of millions in revenue.”

But Zhu believes RL can be much more than a fine-tuning layer. He left Meta to start Pokee, putting RL in charge as the agent’s main decision-maker.

Zhu’s choice to build Pokee outside of Meta came down to one thing: access.

“If Meta were to ask for YouTube’s API, would they ever get it?” he says. “If I were to do it at Meta, [Pokee] would never be internet scale. It would be [within] Meta’s ecosystem.”

No shortcuts

Not just another face in the crowd