2026-07-27 — Your agent just got a legal job description

China now regulates what an agent may decide alone; the EU deferred its high-risk deadlines and still has no agent-specific framework — so build the authority matrix yourself and be compliant in both.

Post

Your AI agent just got a legal job description. China wrote it first.

On 15 July, the Cyberspace Administration's Implementation Opinions took effect — the first binding framework anywhere that regulates the agent, not the model. It splits agent decisions into three tiers of authority:

→ Tier 1 — human only. Signing contracts, moving funds, anything irreversible. → Tier 2 — the agent proposes, the user approves before execution. → Tier 3 — autonomous, but strictly inside an explicitly delegated scope.

The clause engineers keep skipping is Article 6: the boundaries of an agent's authority must be documented BEFORE deployment. Not reconstructed from logs after an incident. And liability follows whoever operated the agent — not whoever trained it.

Meanwhile, Brussels moved the goalposts. Parliament passed the Digital Omnibus on 16 June (423-57, 174 abstentions), the Council signed off on 29 June, and stand-alone high-risk obligations slid to 2 December 2027 — safety components to 2 August 2028. Transparency duties still land on 2 August 2026. And the AI Act still has no agent-specific framework at all.

So two of the three largest AI markets now disagree on the most basic question in agentic engineering: what is an agent allowed to decide by itself?

What I'd build on Monday, regardless of which regulator reaches you first:

Enumerate every irreversible action your agent can physically reach — payments, writes to prod, outbound comms, contract execution.
Put a typed approval gate in front of each one. Not a config flag. A type the compiler enforces.
Log the delegated scope, not just the tool call. "Agent called transfer()" is not evidence. "Agent called transfer() within scope X, delegated by user Y at time Z" is.

That artefact satisfies Beijing today and Brussels in eighteen months.

Deferral is not relief. Harmonised standards land close to the deadline — teams that wait will get weeks to adapt, not months. Building for agent authority now is the cheapest compliance work you'll ever do.

What tier is your agent operating at right now?

#AgenticAI #EUAIAct #AIGovernance #AIEngineering #Compliance

First comment

Sources, for anyone who wants to read the instruments themselves:

CAC Implementation Opinions — analysis with the Article 6 tiers and the "look at control, not code" liability principle: https://blog.pebblous.ai/blog/china-ai-agent-decision-tiers/en/

The CAC original and its official commentary: https://www.cac.gov.cn/2026-05/08/c_1779979789523320.htm https://www.cac.gov.cn/2026-05/08/c_1779983775418216.htm

AI Act Omnibus — the deadline changes in detail (Latham & Watkins): https://www.lw.com/en/insights/ai-act-update-eu-resolves-to-change-rules-and-extend-deadlines

Final adoption status and the confirmed dates (Freshfields): https://www.freshfields.com/en/our-thinking/blogs/technology-quotient/eu-ai-act-unpacked-34-the-final-digital-omnibus-on-ai-key-amendments-to-the-a-102nber

The procedural record, including the 16 June plenary vote (European Parliament): https://www.europarl.europa.eu/legislative-train/package-digital-package/file-digital-omnibus-on-ai

One thing I left out of the post for length: Article 7 also requires verifiable, traceable mechanisms that create permanent behaviour records. The tiers are the headline, but that clause is what turns them into an engineering requirement.

Claims

Sources

Corrections

2026-07-27 — the draft asserted the Omnibus deferral as settled law. The Parliament's own legislative-train record showed the file only "close to adoption" as of 20 June. Verified via Freshfields that the Council gave final approval on 29 June 2026. The post was rewritten to name the procedural chain (Parliament 16 June, 423-57-174; Council 29 June) rather than assert an outcome — more precise, and better positioning. Do not repeat the original phrasing.

Post

Your AI agent just got a legal job description. China wrote it first.

So two of the three largest AI markets now disagree on the most basic question in agentic engineering: what is an agent allowed to decide by itself?

What I'd build on Monday, regardless of which regulator reaches you first:

Enumerate every irreversible action your agent can physically reach — payments, writes to prod, outbound comms, contract execution.

Put a typed approval gate in front of each one. Not a config flag. A type the compiler enforces.

Log the delegated scope, not just the tool call. "Agent called transfer()" is not evidence. "Agent called transfer() within scope X, delegated by user Y at time Z" is.

That artefact satisfies Beijing today and Brussels in eighteen months.

What tier is your agent operating at right now?

#AgenticAI #EUAIAct #AIGovernance #AIEngineering #Compliance

Corrections

2026-07-27 — the draft asserted the Omnibus deferral as settled law. The Parliament's own legislative-train record showed the file only "close to adoption" as of 20 June. Verified via Freshfields that the Council gave final approval on 29 June 2026. The post was rewritten to name the procedural chain (Parliament 16 June, 423-57-174; Council 29 June) rather than assert an outcome — more precise, and better positioning. Do not repeat the original phrasing.