Pokee AI: the agent that clicks the buttons.

It is a Tuesday, and the agent is busy

Somewhere in an enterprise account, a request goes in as a single plain-English sentence: pull last quarter's numbers, build the deck, draft the summary, post it. No macros. No if-this-then-that recipe stitched together by hand. A Pokee agent reads the sentence, plans a route through a dozen different apps, reaches for the right tool at each step, checks its own work, and cites where the numbers came from. Then it stops - because it is finished.

That last part is the whole point. The AI industry spent a few loud years proving that models can write, summarize, and hold a conversation. Pokee AI is built around a quieter, harder question: can a model reliably act? Not narrate the plan - execute it, tool after tool, without fumbling the handoff. In Bellevue, Washington, a team of roughly eighteen people decided that gap was the whole business.

From a Stanford lab to Meta's RL team to a founder's chair

Pokee AI was founded in 2024 by Zheqing (Bill) Zhu, who earned a Stanford PhD in reinforcement learning and then led Meta's applied reinforcement-learning team. Reinforcement learning is the branch of machine learning where a system learns by trial, error, and reward rather than by predicting the next word. It is how software learns to play games at superhuman levels and how robots learn to move. Zhu's wager is that it is also how agents should learn to use tools.

His diagnosis of the current crop of agents is blunt: they are not really intelligent, they are guessing. When an agent picks a tool by asking a language model to "call a function," it is essentially autocompleting its way through a task. That works until the task gets long, the tools get numerous, and one wrong step derails the rest. Pokee's answer is to train the tool-picking itself - to reward the sequences that work and penalize the ones that don't.

One agent, wired into your whole stack

At the center is a foundation tool-use model - a model trained specifically to plan, reason, and select the right tool at the right moment. Pokee says it can extend to more than 6,000 tools and, on function-calling, outperforms GPT-4o, Claude 3.7, and Gemini 2.5 Pro. Around that model sits a platform that connects to 90+ SaaS apps through one-click OAuth, covering well over a thousand integrations, and runs a plan-act-verify-cite loop so the agent shows its work.

A small model punching up

Pokee's marquee open release, PokeeResearch-7B, is a seven-billion-parameter model - tiny by frontier standards. The company reports it leads the field of open 7B deep-research agents, topping 7 of 10 benchmarks. The illustrative bars below show the pitch: state-of-the-art among its weight class, trained with reinforcement learning from AI feedback.

The shape of a seed-stage bet

The $12M seed was led by Point72 Ventures, with participation from Qualcomm Ventures, Samsung NEXT, Salience Capital, SCB 10X, and angels including Typeface founder Abhay Parasnis and semiconductor veteran Lip-Bu Tan. Qualcomm and Samsung are not just money - both have obvious stakes in agents that can run on-device, which is exactly one of Pokee's deployment modes.

Zapier automates rules. LLMs guess. Pokee learns.

To place Pokee, line up the alternatives. Rule-based automation - think Zapier - is reliable but rigid: someone has to build the recipe in advance. LLM-based agents are flexible but brittle: they improvise tool calls and can wander off. Pokee's argument is that reinforcement learning threads the needle - flexible enough to handle a task it has never seen, trained enough to keep picking the right tool.

There is a second wager stacked on the first: privacy. Pokee sells to enterprises with the promise that the agent runs inside their own walls - private cloud, on-premise, or on-device - so sensitive data never has to leave. In a market where "where does my data go?" is the first question a security team asks, "nowhere" is a strong answer.

A short history, moving fast

Things worth knowing

• The name is a job description. "Pokee" hints at what the product does - poking the right buttons across your apps, so you don't have to.
• It runs against the grain. While much of the industry scales ever-larger language models, Pokee trained a smaller model to act.
• It gave away its research model. The enterprise foundation model stays proprietary; PokeeResearch-7B's weights are free under Apache 2.0 - confidence used as strategy.
• Two chip giants are on the cap table. Qualcomm and Samsung both invested - a tell that on-device agents are part of the plan.

Demos and interviews

Find Pokee AI

The same Tuesday, quietly changed

Return to that plain-English request from the start. A year ago it would have been a person's afternoon - hopping between tabs, exporting a spreadsheet, wrestling a template, copy-pasting a summary, remembering to post it. Now it is a sentence, and a Pokee agent working the route while the person does something else. The deck lands. The numbers are cited. The task, unglamorously, gets done.

That is the whole ambition, and it is smaller and stranger than "artificial general intelligence." Pokee AI is not promising a mind. It is promising a worker that can be trusted to take the next correct step, and the one after that, across the ordinary sprawl of software people already use. Eighteen people in Bellevue think that trustworthy action - not clever conversation - is the part of AI still up for grabs. The Tuesday is the proof they are chasing.