IDP to ADP… and other BIG stories from PlatformCon 2026 Live Day London

That was the mantra of PlatformCon 2026 Live Day London, which took place in June. The conference brought together around 1,000 platform practitioners, out of the 280,000 that PlatformCon engages annually across the globe, to discuss the evolution of platform engineering – a discipline designed to deliver infrastructure scalability to large companies, driving growth and efficiency.

Always keen to keep up with the latest developments, Mindera’s platform engineering team were out in force at the event in the UK capital. Unsurprisingly, they found AI was the burning topic. In particular, the development of Agentic Developer Platforms (ADP). The shock was that this was seen to be a key focus for 2027, when in fact Mindera is active in ADP development right now, specifically in Fintech. Nice to see we’re staying ahead of the game!

So, what were the key takeaways from PlatformCon 2026 Live Day London? Attendee and Mindera Head of Platform Engineering Prasanna Krishnamoorthy reveals all while giving his expert opinion on the key topics…

The platform imperative

The common theme across almost every session was that AI doesn’t make platforms less necessary – it makes them more necessary. With developers able to generate code two to 10 times faster using AI*, and agents starting to write, review and deploy, the bottleneck moves from writing code to shipping it safely.

This means it’s the developer platform that decides whether this significant increase in speed is turned into business value or sheer chaos. The recommendation at PlatformCon was not to retire the Internal Developer Platform to ensure the former, but rather to add a new AI layer and turn an IDP into an AI-ready ADP.

Regulation and non-determinism

Platform engineering methods now carry a premium because they are essential to applying AI safely and securely. The big challenges have shifted from “can the agent code?” to trust, governance and economics. And they bite hardest in regulated industries. This reframing was the biggest shift at PlatformCon.

The sharpest single takeaway, however, was the non-determinism reframe. The common fear is that agents are non-deterministic. But that conflates two things. First, the agent’s process and the code it writes are legitimately non-deterministic. Ask it twice and you’ll get two implementations – just like two humans would do. However, what MUST be deterministic is the code’s behaviour.

Of course, this kind of problem is precisely what the standard platform tests, contracts and specifications already check. But with agents being given increasingly critical roles, there is a greater premium placed on these existing engineering practices. In particular, a hybrid method is required: probabilistic for agent-driven tasks and deterministic for pipeline-driven. Taking this approach with an ADP turns a novel governance fear into a familiar, easily managed challenge.

The governance gap and shift to sovereignty

Rather than being bolted onto each agent, governance belongs in the platform. Privilege separation (where agents run least-privilege on an identity that’s always a subset of the invoking human’s) and the validation loop (where failed checks return structured, machine-readable feedback the agent can retry) only work when designed together. These should be cornerstones of every company’s AI policy. The problem is, as highlighted at PlatformCon, that only 70% of organisations have AI policies, yet 89% of platform engineers use AI daily. This creates a gap where regulated firms can get exposed.

Rather than licencing platforms from a SaaS provider, companies are increasingly embracing sovereignty, with economics pushing compute back in-house. Organisations are shifting to self-hosted, open-weights models – which anyone can download, modify, and run – accepting a slight drop in raw intelligence for large gains in privacy and cost control.

Own the boundary, not the model

With sovereignty comes a significant financial outlay on hardware and associated maintenance costs. But this can be softened. At Mindera, we don’t necessarily recommend owning the model. In general, companies only need to own the boundary the model runs inside. Taking an ‘open-weights on bare metal’ approach, where publicly available AI models run directly on an owned platform, delivers 90% of the sovereignty argument at a fraction of the cost and timeline of a frontier programme.

And here's the part that changes where you spend: once you commit to an open-weight model rather than a frontier one, the model is no longer the differentiator. It's a swappable commodity you can re-host or upgrade as better open weights land. So, the money is better invested in the harness, not the model.

The harness – orchestration, context engineering, evaluations, guardrails, the validation loop – is what actually determines whether an open-weight model performs in production, and it's reusable across every model you'll ever run. As highlighted at PlatformCon, around 98% of the value sits in the platform around the model, only 1.6% in the model itself. So, with open weights you stop paying a frontier premium for that 1.6% and pour the budget into the 98% you keep.

Responsibility is converging on ‘liability follows control’

Finally, the PlatformCon roundtable dug into the question no one can dodge: if an autonomous agent does something harmful – say, ships broken infrastructure or places a bad trade – who is responsible? The emerging principle is that whoever controls the system owns the liability. This is consistent with: the EU AI Act Article 14; California AB 316, which now bars “the AI did it autonomously” as a defence; and Singapore’s Jan 2026 framework, which keeps organisations legally accountable for their agents regardless of voluntary compliance. The practical solution lies with building the right accountability structure: designated human supervisors for critical operations, clear override and escalation paths. Plus, full attribution audit trails (agent ID, prompt, model, timestamp) – the same that make “who was responsible” answerable after the fact.

Interestingly, when someone at the roundtable asked who carries liability if a model "decides" to do something dangerous, no one in the room could give a clear answer. What we could agree on was the part that's settled: the vendor controls the weights, but you control the deployment, the permissions, and the guardrails. Plus, the UK’s Financial Conduct Authority doesn't regulate the vendor, it regulates the licence holder. This means the legal liability sits with the authorised firm regardless of how the contract with the vendor is written.

The genuinely unresolved question is what happens when the agent operates at Level 4 autonomy, triggered by a signal with no human in the loop at all. The licence holder is still on the hook, but ‘liability follows control’ starts to strain when no human exercised control in the moment. This was where the room ran out of answers, which is why most firms weren't going past Level 3 yet.

Get your platform AI-ready now

As PlatformCon 2026 London Live Day revealed, companies that want to get the most impact and value from AI need to have a platform that’s fit for purpose.

This means adding an AI layer to your IDP and converting it into an ADP. Or building an ADP from scratch if you don’t currently have an IDP in place.

Don’t wait until 2027 for the predicted ‘year of the ADP’. Instead, gain an edge over your competition and consult with Mindera to get a head start and identify the best path for your organisation.

We hope you enjoyed our round up of trending platform topics from PlatformCon 2026 London Live Day. Get in touch with our platform engineering team today to discuss turning your IDP into an ADP and other platform challenges…