Agon raises £22.5M to train defence AI in synthetic battlefields

Agon, a London defence-AI startup, has launched out of stealth with $30M, or about £22.5M, in funding to build what it calls synthetic battle arenas: simulated environments where autonomous systems and defence hardware are trained against adaptive adversaries. The total is made up of a roughly $7M pre-seed from Lakestar, 201 Ventures, and D3 with angel investors, followed by a $23M seed round led by XYZ, Lux Capital, and Northzone. It is one of the larger early-stage European defence-tech rounds of the year.

The pedigree is notable. Agon is led by co-founder and chief executive Tristam Constant, who served 13 years in the British Army before leading Applied Intuition’s European defence business and holding a senior role at the US defence company Anduril. His co-founder is the serial entrepreneur Junaid Hussain, and the company is chaired by David Helgason, founder of the game-engine company Unity. Agon keeps bases in London and Berlin.

The idea behind the product is that the real world is a poor training ground for autonomous defence systems: rare, dangerous, and impossible to repeat at scale. Agon’s platform instead generates contested virtual environments, drawn in part from live conflict intelligence, in which autonomous systems and the AI that guides them can be pushed against opponents that adapt and fight back. It is pitched at defence companies, AI developers, and the government evaluators who have to decide whether a system is fit for use. The gaming lineage is not incidental: building convincing synthetic worlds is exactly what a game-engine founder knows how to do.

For the UK, it adds to a distinct pattern this year: money flowing into sovereign and defence AI infrastructure rather than only consumer or enterprise software. It follows Valarian’s raise for a sovereign control layer, which we covered recently, and sits alongside the government’s broader push to build up domestic AI capability it considers strategically important. The underlying technique is machine learning, where systems improve from exposure to examples rather than hand-written rules, applied to autonomous agents that must make decisions and act on their own. That framing also carries the obvious weight: training autonomous systems to fight raises hard questions about oversight and accountability that the technology itself does not answer, and which sit largely outside the UK’s current, deliberately light-touch AI regulation. Those are questions worth returning to as this part of the sector grows.