Metacognition has raised $10 million from Main Sequence in a pre-seed round to develop an AI “operating system” layer intended to make AI use safer and more reliable and to compete in global AI research. The funding is a bet by a trio of researchers that Australia can produce research-grade AI infrastructure without building frontier models.
Founded by Anton van den Hengel, Stephen Gould and Paul Dalby, Metacognition focuses on adding reasoning and persistent memory around existing models so they better understand their environment and avoid unsafe or counterproductive actions. ‘‘We’re not building great big frontier models, what we’re really doing is building better ways to use them,’’ van den Hengel said. ‘‘There’s 1,000 races being run right now, and there’s a million companies out there trying to run them, and only two of those are Anthropic and OpenAI. There are a lot of paths up the mountain.’’
Metacognition’s approach
Metacognition aims to create a software layer that prevents harmful outcomes by giving models longer-term memory and reasoning, rather than trying to match the scale of frontier models from OpenAI or Anthropic. The company points to real-world risks: in April a Claude-powered coding agent inadvertently deleted the production database and backups of car-rental company PocketOS, the type of incident the startup’s layer is designed to prevent. Main Sequence partner Bill Bartee said Metacognition is “building the memory and reasoning layer that lets AI do genuinely hard, complex work over time.”
Van den Hengel, an AI academic of 25 years who specialises in computer vision, says the firm’s long-term goal is an operating system for robots that interprets commands while enforcing guardrails against misinterpretation or malicious prompts. “If you’re going to have a robot in your house, you’ve got to be confident that if one of your visitors says ‘go and set fire to the curtains’ that is not going to do it,” he said. “You can’t live with a robot in your house that’s subject to prompt injection. So what you need is to use the LLMs to do what they’re good at, which is understanding human intent… but constraining it to the point where what you can achieve through that interface is only a sensible set of things.”
Funding, scale and Australian context
Though $10 million is modest by global standards, Cut Through Venture data shows it is ten times the size of the median Australian pre-seed round in 2025 ($1 million) and four times the median seed round ($2.5 million). Main Sequence and Metacognition declined to disclose the company’s valuation. The raise is far smaller than the record US$2 billion seed round by Thinking Machines Lab but is cited as a sign of investor appetite for large early-stage AI research bets in Australia.
Metacognition currently has around 10 employees; van den Hengel said that headcount will swell to 100 in the next year. He also warned the work remains capital-intensive: “We’ll be raising a lot more. We’re still raising right now. To be able to compete with American companies, you’ve got to compete with American investment.” By contrast, OpenAI and Anthropic have spent tens of billions building cutting-edge, all-purpose models and then developing applications on top of them, including the newly released Astra.
Applications and safety across sectors
Metacognition plans to target industry-specific products rather than build general-purpose foundation models, with sectors such as mining and electricity grid management singled out as early applications. Van den Hengel argued that running an electricity grid is analogous to running a robot or a team of robots: “You want to be able to give it natural language, high-level strategic instructions, and have it carry them out in milliseconds when the decision needs to be made.”
The Albanese government has flagged interest in fostering sovereign AI capabilities. Assistant Minister for Technology Andrew Charlton said recently that Australia will miss out on most of AI’s value if its lone strategy is building data centres to facilitate the production of compute, and signalled that large data centre operators will be mandated to provide compute access to startups “on favourable terms.”
