Product1 publisher2 min readPublished
Four of seven AGNTCon vendors sell agent governance built on their pre-AI products
SiliconANGLE's account of AGNTCon + MCPCon Europe in Amsterdam describes gateways, an AI jail, deterministic orchestration and an EU AI Act catalog. The frontier model makers stayed off the show floor.
The Product Desk · Product desk
What happened
- AGNTCon + MCPCon Europe 2026 ran in Amsterdam in the week before September 19, when SiliconANGLE published its walk-through of the exhibit hall and the vendors selling into agentic AI.
- None of the frontier model vendors appeared on the show floor, according to the report, leaving the hall to startups selling the infrastructure that sits around agents.
- Traefik Labs showed API, AI and MCP gateways as three products sharing one binary, pitched as a control plane producing evidence of agent behaviour fit for regulatory, audit and forensic use.
- Bluerock Security's governed workspace, which it calls an AI jail, controls which models a desktop coding agent can reach, manages its token usage and integrates with enterprise identity systems.
- Grape Up's Aiboostr carries a built-in catalog of each agent's owner, purpose and users, the inventory the EU AI Act requires organisations to maintain.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision A pilot can be scored on what the vendor produces at the end, not on whether the agent finished the task: the permission boundary it ran inside and a log an outside auditor can read. Both are testable in thirty days without waiting on a roadmap item.
- constraint Any agent a developer builds for European use needs a register entry with a named owner, so the rollout budget has to cover keeping that inventory current, not only the platform licence.
- exposure Model choice and token spend currently sit with whoever installed the coding agent on their machine; a governed workspace hands both to whoever administers identity, and that person then answers for the bill.
- contradiction The report counts paying customers on the floor, but for the four governance products it gives no customer and no price, so it cannot show whether buyers are paying for constraint or merely approving of it.
A developer running a coding agent on a desktop currently decides which model it calls, how many tokens it spends and which files it opens. Bluerock Security sells a governed workspace it calls an "AI jail" that moves those decisions behind enterprise identity management [9]. The isolation layer underneath it is a military-grade microkernel hypervisor the company built to keep data separated on bare metal [8]. Bluerock has also extended static application security testing to MCP servers and agent skills [10].
All four of the vendors the SiliconANGLE account files under governance and orchestration extend a product that predates agents [16]. Traefik Labs is selling an API, AI and MCP gateway as three products sharing one binary, on top of its cloud native proxy and ingress controller [6]. Orkes came out of Netflix as a scalable cloud native orchestration platform and now sells deterministic agent orchestration that operators can inspect and track [11]. Grape Up built Aiboostr, its orchestration and governance platform, on its Europe-based cloud and data management work [12].
Four of the seven companies the writer picked as representative of the show sit in those two sections [15][3]. Traefik describes the target as a unified control plane that delivers evidence of agentic behaviour detailed enough for regulatory, auditing and forensic requirements [7]. "Governance may be the root canal of information technology," the SiliconANGLE report said [14].
The report says the floor had plenty of bona fide paying customer stories, and also that innovation is outrunning adoption in a market where the real products are sometimes hard to separate from the hand-waving [4][5]. It does not name a customer or a price for any of the four [17]. So the exhibit hall evidence covers what vendors chose to build, and the question of whether permission boundaries and audit trails are the part buyers actually pay for stays open.
For a pilot, the useful test is what the vendor can hand over on the last day:
- the permission boundary the agent ran inside, written down before the run; - a log an auditor outside your team can read without a walkthrough from the engineer who built the thing; - a register entry naming the agent's owner, purpose and users, as the EU AI Act requires organisations to track [13].
That third item is the one teams discover late, because it is clerical and a demo runs fine without it. Grape Up points at Aiboostr's built-in catalog to produce it [13].
What to watch
- Whether Traefik, Bluerock, Orkes or Grape Up publish a named customer and a price for the governance tier.
- Whether frontier model vendors take exhibit space at the next AGNTCon and ship their own permission and audit surfaces.
- Whether EU AI Act inventory duties turn up as a scored line item in agent procurement documents.