Metacognition AI's 'Windows for robots' pitch lands $10M from Main Sequence

Metacognition AI, a newly founded startup building what it calls "Windows for robots," has raised $10M from Main Sequence, the deep tech venture fund backed by CSIRO. Capital Brief first reported the raise.

The company was founded this year by AI researcher Anton van den Hengel, chief scientist of the Australian Institute for Machine Learning, together with Stephen Gould and Paul Dalby.

At the centre of the pitch is an operating system for AI agents that sits underneath large language models, controlling how those models interact with machinery and other systems. The promise is that robot hardware can be trained to do useful tasks by non-experts using simple verbal instructions and feedback — no specialist programming required. Robots would work alongside humans, remembering instructions and improving their performance as they interact with their environment, with users teaching the agents specific tasks as they go.

Metacognition says it "will transform the AI revolution, leading to more capable and safer AI agents that are more personalised to their user," and describes its platform as "the missing software platform that finally enables robot hardware to be trained to do useful tasks by non-experts with simple verbal instructions and feedback."

Rather than chasing consumer robotics, the startup is initially targeting existing industrial systems. It has built simulations in which operators can direct an open-cut mine or manage an electricity grid simply by speaking to the system — a demonstration of how natural-language control might work in environments where the machinery is already installed and the expertise is held by workers rather than engineers.

The company currently runs a team of eight and is expanding, hiring AI researchers. The raise follows a pattern of deep tech investment flowing into the intersection of robotics and foundation models, where the constraint is increasingly software — how machines are instructed and supervised — rather than the hardware itself.