There is no Article 12 without the runtime

Article 12 of the EU AI Act becomes enforceable for high-risk AI systems on 2 August 2026. It puts the obligation to keep logs on the deployer. It cannot be met unless the AI developer emits them. Until the runtime layer adopts the open hook standard at agenthook.org, every regulated EU enterprise running an autonomous agent will be non-compliant on day one of the enforcement window.

Article 12 of the EU AI Act is short and unambiguous. High-risk AI systems "shall technically allow for the automatic recording of events" over the system's lifetime, in a way that ensures traceability for the deployer, the auditor, and the regulator.

Read that again. The technical capability sits with the provider. The recording obligation sits with the deployer. If the runtime does not emit the events, the deployer has nothing to record, and Article 12 is unmeetable. No amount of enterprise governance closes that gap. You cannot log what was never offered up.

The compliance chain breaks at the runtime

Every major agent vendor emits something. Claude Code emits hook events. Cursor emits its own format. Amp emits another. The OpenAI Agents SDK emits a third. Each shape is different. Each names the events differently. Each carries different reasoning context, or none at all.

A regulated EU bank running five different agents from five different vendors ends up with five different audit trails, in five different shapes, none of which a regulator can correlate. That is not Article 12 compliance. That is five vendor-specific log streams in a trench coat.

The fragmentation is not anyone's fault. There has never been an agreed surface to emit. Every runtime team had to invent one. The result is what you would expect: a working capability, expressed five different ways, with no shared vocabulary the deployer or the regulator can read across.

Ten events. One wire format. One open spec.

agenthook.org defines the minimum surface every conforming runtime should emit:

• PreToolUse — before tool execution
• PostToolUse — after tool execution
• UserPromptSubmit — user submits a prompt
• PreLLMCall — before the LLM call
• PostLLMCall — after the LLM call, carrying reasoning and response
• ModelResponse — model finishes generating
• SessionStart — new session begins
• SessionEnd — session ends
• AgentHandoff — an agent delegates to another
• ErrorOccurred — non-terminating error

One envelope. One canonical event vocabulary. One thing for auditors to read.

It is deliberately small. It is deliberately runtime-agnostic. It is deliberately open. The point of a standard is that no single vendor owns it, and every vendor can ship it without giving up product surface area.

This cannot be a vendor feature

A single vendor shipping its own logging format does not make EU enterprises compliant. It makes them compliant with that vendor. Stack four vendors and you have four compliance silos and zero cross-agent traceability.

Article 12 is a property of the deployment, not the product. A property of the deployment can only be guaranteed by a standard the entire deployment honours. That is why the open hook surface has to be adopted upstream, by the runtime providers, before any regulated enterprise can honestly claim Article 12 readiness.

This is the same shape as TLS, OAuth, OpenTelemetry. None of those are vendor features. All of them are standards the entire layer agreed to honour. The runtime layer of agentic AI is the next one in the queue.

What we are asking the runtime providers to do

Implement the agenthook.org envelope. Emit the ten canonical events. Publish to a hook bus the deployer controls.

That is the entire ask. There is a reference implementation, HookBus, that any vendor can subscribe to today. There is an open spec at agenthook.org. There is a deadline.

The deadline

Article 12 becomes enforceable for high-risk AI systems on 2 August 2026. That is roughly fourteen weeks from the date of this post. The product decisions that determine whether a given runtime is compliant on that day are being locked into roadmaps this quarter. Runtimes that emit a clean, standard, agenthook-conformant event stream will be procured. Runtimes that emit a vendor-specific log dump will face procurement objections from every regulated buyer in the bloc.

The Commission has proposed a delay through the Digital Omnibus package, with trilogues underway as of March 2026. Nothing has passed into law. 2 August 2026 remains the enforceable date, and procurement teams in regulated industries are not waiting on uncertain legislative outcomes to make compliance decisions. They are asking the question now.

What to do now

If you build an agent runtime: implement agenthook.org. The spec is open. The reference implementation is open. There is no licensing barrier and no commercial lock-in. You join a standard. You do not buy into a product.

If you deploy AI agents in the EU: ask every vendor in your estate, in writing, whether they will emit agenthook-conformant events. The answer is your compliance position. Make it part of every renewal and every new procurement.

If you advise on AI governance: the runtime layer is the load-bearing wall. Policy frameworks that assume the events will appear are not policy frameworks. They are wishlists.

Article 12 cannot be solved at the deployer. It has to be solved at the runtime. agenthook.org is the open standard that scales beyond one vendor.

Join the working group   Read the spec

If you would prefer a private walkthrough of HookBus, the reference implementation, book a demo.

← All posts