Australian AI research startup Metacognition AI has secured a AUD $10 million pre-seed investment from Main Sequence, the deep tech venture capital firm backed by CSIRO. The company was founded by artificial intelligence researcher Anton van den Hengel, fellow researcher Stephen Gould, and research strategist Paul Dalby. It is building a Windows-like software layer that allows people to safely control robots and complex industrial systems using natural language.
An Operating System Beneath Large Language Models
The startup's first product, Zeus, is described as an operating system for agents that sits beneath large language models. It controls how those models interact with machines, software, and other systems while adding reasoning and persistent memory. This design gives businesses more control over AI behaviour and reduces the risk of unsafe or counterproductive actions when models are applied to real problems.
Investor Confidence in a Different AI Approach
Main Sequence partner Bill Bartee said every model is limited by its context window, and Metacognition is building the memory and reasoning layer that lets AI do complex work over time. The $10 million raise is ten times the median Australian pre-seed round in 2025 and four times the median seed round, according to Cut Through Venture data. The parties declined to disclose the valuation, but the deal signals growing appetite for ambitious early-stage AI research in Australia.
From Household Robots to Industrial Control
Metacognition's long-term vision centres on robots that understand verbal commands while operating within strict guardrails. Van den Hengel noted that a household robot must not obey harmful instructions from visitors, meaning large language models need constraints that limit actions to a sensible set. The company is first targeting existing industrial systems, with simulations where operators direct an open-cut mine or manage an electricity grid through natural language.
Scaling a Deep Tech Team
Metacognition currently employs around 10 people and plans to expand to approximately 100 within the next year. Van den Hengel said the company is still raising capital because competing with American AI companies requires competing with American investment. Building harness technology is cheaper than training frontier models, but the work remains expensive and will demand significantly more funding over time.
A Focused Path in a Crowded Market
Unlike OpenAI and Anthropic, Metacognition is not building frontier models but better ways to use them. Van den Hengel observed that many paths exist in the artificial intelligence market, and only a small number of companies are racing to build the largest models. The startup will target specific sectors such as mining and build products from the ground up, aiming for AI systems that users can trust with sensitive access such as bank accounts.
Metacognition's raise highlights the potential for Australian research teams to pursue significant AI infrastructure opportunities without building frontier models. By placing safety, memory, and reasoning beneath large language models, the company aims to make advanced AI practical for industrial systems and eventually household robots. Its early focus on real-world control problems positions it as a notable player in the emerging agent software market.
Source: CapitalBrief