Why Multi-Agent Architecture Matters: Learnings from the FCA AI Supercharged Sandbox

There’s no shortage of conversation about AI in financial services.

Far less about how it should actually be built when decisions are high-stakes, regulated, and affect real people.

That’s what made the Financial Conduct Authority AI Supercharged Sandbox such a valuable programme for us at Serene. Not as a demo opportunity, but as a genuinely explorative environment to test architecture, guardrails, and judgement.

At its core, the challenge we’re focused on is simple to state, but hard to execute:

build for scale, while keeping people firmly at the centre.

(Which is usually the moment a seasoned developer sharpens their pencil and reaches for their notepad.)

An environment designed for responsible exploration

Before getting into the technology, it’s worth calling out the environment itself.

Huge credit to Matt Lowe, Colin Payne, Séamus Merrin and the wider FCA Innovation team for creating a space that was calm, rigorous, and genuinely supportive of responsible exploration. That tone really matters, as it shapes the quality of thinking that follows.

We also benefited from engaging with multiple mentors throughout the programme, each bringing a different lens on architecture, regulation, delivery, and scale. That breadth of challenge helped sharpen our thinking.

Special shout-outs in particular to David Tracy and his team, Navneet Mathur, Rémi CUCHILLO, and Karen Elliott, whose thoughtful challenge pushed us to translate AI architecture into something that could deliver real, frontline impact.

Why we’re focused on multi-agent architecture

Rather than relying on a single monolithic model, we’ve been exploring how multi-agent AI can act as an enabling layer over our core platform.

One of the key advantages of this approach is flexibility. Multi-agent architecture allows firms to adopt and combine capabilities selectively, based on their operational context and maturity rather than forcing everything through an all-or-nothing system.

In practice, that means:

• Different agents interpreting behavioural and transactional signals
• Separate agents assessing severity, confidence, and prioritisation
• Orchestration logic determining when insight should surface, and when it shouldn’t

This separation creates clear boundaries between insight, recommendation, and action, preserving human judgement while improving signal quality upstream.

What the Sandbox allowed us to test

The real value of the Sandbox was seeing how this orchestration behaves under real operational constraints - where clarity, timing, and ability to act matter just as much as detection accuracy.

In particular, we explored how multi-agent orchestration can translate our risk detection (SereneID), harm prediction (SereneScore), and support recommendations (SereneCare) into real-time, contextual guidance delivered directly within existing operational systems and workflows - from CRMs and agent desktops to chatbots and AI co-companions - without overwhelming teams or removing human judgement.

That included testing:

• How early risk signals surface early enough, and close enough to the workflow, to be useful
• How insight becomes proportionate guidance, not abstract scores

Where explainability genuinely helps teams understand why something surfaced

• How tone, timing, and next steps adapt to the interaction context

This is where architecture stops being theoretical and starts shaping advisor confidence, escalation accuracy, and real customer outcomes.

A standout moment

One highlight (a tech fan-boy moment!) was the session from Professor James Fergusson on AI hype versus reality.

His walkthrough of agent composition and orchestration strongly reinforced the direction we’ve taken with MySerene: that complex, real-world problems are best handled by systems of agents with clear roles, rather than a single model trying to do everything.

What resonated most was the emphasis on combining agents to interpret signals, reason over context, and iterate through learning loops. Particularly when that approach enabled his team to secure all ten of the top ten places in an official, human, paper-writing competition!

It was one of those moments where you quietly nod, scribble a few notes, and think: good, we’re building in the right direction.

I knew I liked this James for a reason. (It wasn’t just the beard admiration.)

Looking ahead

The Supercharged Sandbox has given us valuable ideas and plenty of food for thought, insights that are now shaping our product roadmap.

As we take these learnings forward, our focus remains the same: strengthening client value responsibly, through technology and insights that teams can trust, understand, and stand behind.

And yes the view from the 15th floor is marvellous!

More to come.

This post was originally published on LinkedIn. Read the original here.