PlatformCon Paris 2026: Rethinking Platform Engineering for AI Agents

This article was originally published in French on the Seven Mile blog.

Last week, on September 24, I had the opportunity to attend PlatformCon Paris on behalf of Seven Mile. The event was organized by WeScale and featured around twenty sessions. Unsurprisingly, the main theme was AI and governance applied to platforms: agents, security and automation.

In the opening session, Luca Galante explained that one team in ten sees the model as its main problem, while around 30% point to the infrastructure around it. Since everyone has access to the same models, it’s now the platform that really makes the difference.

In another session, Alexandre Hugla (Broadcom) drew a consequence from this: agents must now be treated as a full-fledged type of user, with tracking, auditing and access control applied to them.

From that day, I mainly kept four talks in mind, and I’d like to go back over them with my own take.

Manuel Pais: an agent doesn’t see the boundaries between teams

Manuel Pais, co-author of Team Topologies, opened his talk with an anecdote. An agent tasked with unblocking a customer report uses a data endpoint owned by another team, which was never meant for a third-party user. Nobody checks behind it, and three weeks later the problem comes back with even more customers impacted.

The thing is, the agent did its job, but it had no awareness of all the organizational boundaries you find in a company. His point is that organizations now need to rework machine-readable APIs that tell the agent what it can and cannot do, so that it doesn’t cross those boundaries. It’s also a way to optimize the agent’s cognitive load, because the context window is in fact its working memory.

He also says that AI can help reconcile autonomy and accountability. At Seven Mile, we partly agree. We’re fully convinced that AI can produce the evidence, in the form of traces, policy checks and continuously verified compliance. Accountability, however, must always stay human.

As you can see on the slide below, agents don’t know the business intent: it’s up to people to stay in the loop. What ultimately lets the two coexist is observability. You can’t delegate autonomy to an agent you don’t know how to observe.

Pablo Lopez: every agent is an identity

Pablo Lopez, who leads the Platform Engineering teams at PayFit, talked about two incidents he went through recently:

1. the compromise of the axios npm package at the end of March 2026;
2. the theft of a GitHub token, reused a few months later in connection with the Miasma worm.

He explained that at PayFit, signed commits were spotted because of their timestamp: 4 a.m. Their response comes in three steps:

1. Know, by setting up threat watch. It can be automated: RSS feeds from security blogs go through an LLM, which tells you which compromised packages affect what runs on the platform. You also need SBOMs that come back in 5 minutes instead of several hours.
2. Act, with kill switches: emergency mechanisms that let you immediately lock a compromised workstation, revoke all its tokens or block merges. During the second incident, merges were blocked for 8 hours.
3. Prevent, by applying least privilege, going deny by default and giving its own identity to every service, every agent and every sub-agent.

We agree with the principle, but the tooling isn’t mature yet. For the inventory of non-human identities, we’re already behind on the tokens that humans create. Pablo Lopez adds the secrets created by the agents themselves, and with ephemeral sub-agents, this can’t be managed by hand. You need short-lived identities, issued by the platform at the moment the agent needs them. This really is a capability that should be provided as a shared service, rather than leaving each team to manage its own access.

Michel Hubert: the control plane and graduated delegation

Michel Hubert (Avanade) starts his talk from this observation: an agent triggers a chain of executions whose latency, cost and results vary. A pod’s CPU can sit at 5% while the user complains, because the pod is waiting on the model or on a tool. The platform contract therefore shifts from resources to outcomes under constraints, and the platform becomes a control plane that decides which rules apply.

His conclusion is that we now need graduated delegation, for example:

• restarting a pod: the agent does it alone;
• shifting 5% of traffic: with verification;
• deleting data: with human approval.

This graduated delegation is also what we apply to our own agents. We separate the actions that can be automatic, left to the agent, from the more critical ones that require a human in the loop. A simple example: an agent that sorts a mailbox can apply labels and archive on its own. On the other hand, an agent that prepares an email must stop at the draft: sending remains a human decision. These rules can be written into the instructions the agents read, for instance in the skills they use.

At the scale of one person, it looks handy and easy, but at the scale of a company, that’s exactly the role of this control plane.

Benjamin Ameztoy: opening observability to agents

Benjamin Ameztoy, Platform Tech Lead Observability at Peaksys, Cdiscount’s tech subsidiary, told the story of five years of in-house observability: 5,000 machines, mostly on-premise, around 2,000 microservices, and a deployment every 7 minutes on average. Metrics, logs and traces live on three separate platforms behind Grafana, to keep visibility if one goes down. The current project: opening this data to AI agents through APIs and MCP, while keeping an eye on platform stability under the unoptimized queries agents generate. After five years, the hardest part, according to him, is still convincing a hundred teams to structure their logs and instrument their applications.

Opening observability to agents through MCP is the right direction, and that’s where graduated delegation applies: the agent reads, with query quotas, and doesn’t act without approval. Benjamin Ameztoy’s concern about unoptimized queries shows that agent experience is also designed on the side of the platform that absorbs those queries.

Rethinking the job

Today, the platform engineer’s job needs to be rethought. Platform teams must become able to design platforms, and the governance around them, that can host a new type of user: AI agents.

In his closing words, Luc Legardeur, CEO of WeScale, pointed out that the engineer’s role is moving from execution to design and governance, because writing code, deploying clusters and delivering environments will be largely automated by these agents. In his talk, Michel Hubert suggested moving from operating an infrastructure to designing the rules by which it should operate itself.

For a platform team, the first step is quite concrete. Go through each capability of the platform and write down:

• what an agent can do alone;
• what it can do with verification;
• what always requires human approval.