What actually happened in Amsterdam
According to SiliconANGLE's report from the show floor, the agenda was dominated by enterprise deployment mechanics rather than existential risk. That is not because the risk debate went away — it escalated. The conference simply moved somewhere else: to the plumbing.
The most telling detail is what was missing. The frontier model vendors — OpenAI, Anthropic, Google DeepMind — were not on the exhibit floor, according to the same report. That left the space to open-source maintainers and startups shipping production agent applications.
Anyone who has run an agent in production knows why. The hard part was never the model. It is the identity layer: which agent, acting on whose behalf, with which scopes, writing to which system, and can you reconstruct that six months later when an auditor asks. On our own AI Arena rig we can spin up a local model on an RTX 5060 Ti and get an agent loop running in an afternoon. Making that loop accountable is a project, not an afternoon.
From observability to provable control
The phrase doing the rounds in Amsterdam was "provable control" — the idea that governance should not just log agent behaviour but cryptographically verify authorisation and action in real time. That is a meaningful jump from the dashboards most teams deployed in 2025. Observability tells you the agent did something odd. Provable control is supposed to stop it before the write call leaves the building.
The concrete example from the floor: Traefik Labs has collapsed three separate products — API Gateway, AI Gateway and MCP Gateway — into a single binary, according to the SiliconANGLE report, positioning it as a sovereign control plane. One binary means one place where policy is enforced and one place where the audit trail lives. From an operator's perspective, that is worth more than another dashboard.
The transport layer underneath is also stabilising. The Model Context Protocol's maintainers have published updates including stateless transport and enhanced cybersecurity metadata — the kind of unglamorous spec work that decides whether your gateway can be horizontally scaled behind a load balancer without sticky-session pain. The spec is public at modelcontextprotocol.io if you want to check whether your existing server implementation is affected.
The money is chasing auditability, not safety essays
Venture capital followed the same logic. Agent certification startup AIUC said at the event that it has raised $40 million to audit frontier models and — crucially — to unblock enterprise security reviews that stall deployments for months, according to SiliconANGLE's report. Certification is a boring business until it is the thing standing between your agent and a signed contract.
One estimate cited at the conference and included in the SiliconANGLE report deserves more attention: automated agentic workflows and bots are estimated to generate 20% to 30% of content in corporate brand-monitoring workspaces. That is not a future scenario. That is roughly a third of what some teams are reading, produced by software nobody has yet certified. Whatever you think of agent risk, the measurement problem is already here.
The European timeline is not a future event
Here is where the European read of this conference differs from the American one. Since 2 August 2026, the EU AI Act's Article 50 transparency obligations have applied, with enforcement by the European AI Office and national competent authorities. That does not mean all deployer obligations became binding on that date: high-risk-system obligations follow the Act's staggered timetable, with different compliance dates depending on whether the system falls under Annex III or Annex I product-safety legislation. For teams deploying agents in the EU, the question is no longer "should we prepare" but "can we produce the documentation today".
That reframes the Amsterdam vendor hall considerably. An agent control plane that emits verifiable authorisation records is not a nice-to-have in the EU — it is closer to a compliance artefact. And it explains why open-weight sovereignty has moved from ideology to procurement line item.
What agents actually cost to run right now
Nobody at a conference hands out an honest cost sheet. A reliable per-model price comparison needs current official vendor pages, and several of the specific rates mentioned on the floor could not be verified against official documentation as of this week. For that reason, we are not publishing a table of per-million-token prices here. What a European buyer can verify is the structure of the costs: output tokens dominate agent budgets, because every reasoning step writes more tokens than it reads; long-context tiers can sharply increase the bill once a conversation crosses a vendor's context boundary; and open-weight models carry infrastructure costs instead of API margins, but remove per-token vendor lock-in.
European availability has also stopped being the weak spot. Several EU providers now offer in-region endpoints with contractual data residency, custom rate limits and SLA-backed inference — the kind of paper trail that satisfies a procurement review that a global API endpoint does not. That compliance-shaped demand is now a procurement line item in Europe.
What I would actually do with this
If you are running agents in production in the EU today, the order of work is roughly: put every agent behind one gateway so identity and policy live in one place; make sure your MCP servers are compatible with the stateless transport changes; and start logging authorisation decisions in a form you could hand to an auditor, not just a form you can grep. Certification vendors like AIUC exist precisely because enterprise security reviews are the bottleneck, so check whether a certificate shortens your sales cycle before assuming you need to build everything yourself.
The extinction debate is not going away. But the industry that showed up in Amsterdam has made its choice: it is treating agents as regulated infrastructure. In Europe, that choice was made for it about six weeks ago.
What is a "sovereign control plane" in plain terms?
A single enforcement point — usually a gateway — where agent identity, permissions and audit logging are applied, with data and policy kept inside a jurisdiction you can name. The Traefik example is a single binary covering API, AI and MCP traffic. The point is that policy is enforced in one place instead of scattered across five services.
Should a small EU team pick open weights or a US API for agent workloads?
It depends on volume and sensitivity. For high-volume, low-complexity tasks the permissive open-weight tier — MIT or Apache 2.0 — can be dramatically cheaper and removes the data residency question entirely, at the cost of GPU hardware or EU hosting fees. For complex reasoning, frontier APIs still lead, and the practical question becomes whether the vendor offers an in-region EU endpoint with contractual residency guarantees.
Do this week's MCP transport changes break existing servers?
Stateless transport is an addition to the spec rather than a forced migration, but gateways and load balancers configured around sticky sessions are the parts most likely to need rework. Check your gateway vendor's compatibility notes before you scale horizontally.
