AI Blog

China Wrote Down the Agent Design Doc Everyone Skipped

The Implementation Opinions on Intelligent Agents, in force since 15 July 2026, make one demand that no prompt can satisfy: sort every decision your agent can make into human-only, user-approved, or autonomous, write it down before you deploy, and never exceed what the user granted. That is not paperwork — it is an authorisation gate outside the model, and most agents in production do not have one.

By Agentic AI Wiki 12 min read

Somewhere in your agent's system prompt is a sentence like "always ask the user before making a purchase." That sentence is the entire autonomy boundary of most agents shipping today, and as of 15 July 2026 there is a national regulator that considers it insufficient by construction. China's Implementation Opinions on Intelligent Agents require every deployed agent to sort its decisions into three tiers — human-only, user-approved, autonomous — before deployment, and to never exceed the scope the user granted. You cannot satisfy that with a prompt. You satisfy it with a gate outside the model, which is exactly the thing good agent architecture was already converging on.

At a glance

One document, three ministries, and the first time a national government has treated agents as a regulated category in their own right rather than as an application of generative AI.

ItemDetail
DocumentImplementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents (智能体规范应用与创新发展实施意见)
Issued byCyberspace Administration of China, National Development and Reform Commission, Ministry of Industry and Information Technology — jointly
Published / in force8 May 2026 / 15 July 2026
Instrument typeImplementation Opinions — a policy directive, not a statute. Sets required practice and signals what codifies next.
Core obligationA documented, three-tier decision-authority boundary per agent, established before deployment
Scope of applicationClassified and tiered by scenario: sensitive sectors carry filing, testing and recall duties; low-risk uses rely on self-assessment and industry self-discipline
Obligations by application class under the Implementation Opinions A three-by-four grid showing how obligations scale with application class. Sensitive sectors such as healthcare, transport, media and public safety carry filing, compliance testing and recall provisions; general commercial deployments carry documented authorisation boundaries and self-testing; low-risk uses such as entertainment and office work rely on self-assessment, reporting and industry self-discipline. The documented three-tier authorisation boundary is required in every class. How the obligations scale with application class Documented tier boundary Filing & compliance testing Recall provisions Who checks Sensitive sectors health, transport, media Required before deployment Mandatory Yes Regulator + sector body General commercial most enterprise agents Required before deployment Self-testing, reporting Not specified You Low-risk uses entertainment, office Required before deployment Self-assessment tooling No Industry self-discipline Applies in full Lighter form Not imposed
The heavy obligations are scoped to sensitive sectors — but the documented authorisation boundary is the row that does not vary.

The rule that actually bites

Most of the document is what you would expect from an industrial policy: a definition of intelligent agents as systems capable of autonomous perception, memory, decision-making, interaction and execution; nineteen named application scenarios spanning scientific research, industrial development, consumption, public welfare and social governance; commitments on standards bodies, open-source ecosystems and evaluation tooling. It is written in the register of 放得活 and 管得好 — keep it lively, keep it governed — and most of it is promotion rather than restriction.

Then there is the authorisation requirement, and it is unusually specific for a document of this type. Before an agent is deployed, the boundary between three classes of decision has to be established and written down:

  • Decisions only the user may make. The agent may prepare, recommend and explain; it may not act, and no delegation makes it able to.
  • Decisions the agent may execute only after user authorisation. Approval is per-action or per-scope and precedes execution.
  • Matters the agent may handle autonomously, strictly within a delegated scope.

Attached to that are two clauses which are easy to skim past and which do most of the work: users retain the right to be informed of the agent's autonomous decisions and the final say over them, and the agent's execution must not exceed the scope the user approved. Read them as engineering requirements rather than as principles and it becomes clear what is being asked for.

Why a prompt cannot satisfy this

Where the autonomy boundary has to live Two stacked architectures. In the top one the autonomy rule sits inside the system prompt, so the model both proposes and authorises the action and attacker-supplied context can move the boundary. In the bottom one the model only proposes; a separate authorisation gate checks the action against a recorded scope grant and either executes it, escalates it for user approval, or refuses it, writing every outcome to an audit log. Boundary in the prompt — unenforceable Context incl. untrusted text Model decides AND authorises Tool call executed World irreversible One phrasing in the context moves the boundary Boundary in the gate — what the three-tier rule requires Context incl. untrusted text Model proposes only Authorisation gate action × scope grant outside the model Tier 1 — refuse human-only decision Tier 2 — escalate user approves first Tier 3 — execute within delegated scope Audit log — proposal, tier, grant, outcome append-only, not writable by the agent
The same three tiers, implemented in the two places they can live. Only the lower one is checkable.

An instruction is not a boundary

If the rule "ask before purchasing" lives in the system prompt, then the component enforcing the boundary is the same component that reads attacker-controllable text, that is sampled stochastically, and that will occasionally decide this particular purchase is obviously what the user wanted. It is not a control; it is a strong suggestion made to a probabilistic system. The clause "must not exceed the scope the user approved" is unsatisfiable in that design, because nothing in the system knows what the approved scope is in a form anything can check.

Move the boundary outside the model and every clause becomes implementable. The model proposes an action; a gate reads the action's type and parameters, looks up the scope the user actually granted, and refuses, escalates or executes. The tier is a property of the action, resolved by code. "Never exceed the approved scope" becomes a comparison between two data structures instead of an aspiration.

The three tiers are a permission model you already needed

Written out, tier 1 is an action the agent's credentials should not be able to perform at all; tier 2 is an action requiring a fresh authorisation token; tier 3 is an action inside a standing grant. That is ordinary authorisation design, and it has a well-understood implementation: scope the credential rather than the prompt, keep a record of what was granted and when, and make the grant expire. The regulation's contribution is not the mechanism — it is the requirement that you enumerate the actions and classify them before deployment rather than discovering the classification during an incident.

Anyone who has built agents seriously will recognise that enumeration as the design document they meant to write and did not. It is tedious, it forces arguments about which actions are actually reversible, and it is the single most useful artifact an agent team can produce. A regulator has now made it mandatory in one large market, which is an unusual way for a best practice to spread but not an ineffective one.

The clause about being informed is a logging requirement

"Users retain the right to be informed of autonomous decisions" cannot be met by a system that does not record what it decided autonomously and why. In practice this means a per-action record — proposal, tier, the grant relied upon, the outcome — held somewhere the agent itself cannot rewrite. Teams that already run agent tracing have most of this; teams that log at the conversation level have almost none of it, because the unit of record is wrong.

How this differs from the frameworks you already track

Three ways regulators cut the same problem Three columns comparing what each framework classifies. The EU AI Act tiers whole systems by application risk, producing one label per product. China's Implementation Opinions tier individual decisions inside a system, producing a table of actions. The NIST AI Risk Management Framework tiers nothing and instead describes a process an organisation runs. What gets classified EU AI Act Tiers the system by application risk Output: one label for the product CN Implementation Opinions Tiers each decision inside the system Output: a table of actions and grants NIST AI RMF Tiers nothing; describes a process Output: a programme the org runs Only the middle one names a control surface that lives inside the agent loop.
Same problem, three different units of analysis — and only one of them lands inside the agent loop.

The interesting difference is not strictness, it is the unit of classification. The EU AI Act tiers systems: you determine whether your product is prohibited, high-risk, limited-risk or minimal-risk, and a set of obligations attaches to the product. The Implementation Opinions tier decisions: one agent contains actions in all three tiers simultaneously, and the classification is a table, not a label. The NIST AI RMF, meanwhile, tiers nothing and describes a process an organisation runs.

Decision-level tiering maps onto agent architecture in a way system-level tiering does not. An agent is not one risk level — a coding agent that reads files, runs tests and opens pull requests spans three obligation classes in a single loop, and a product label cannot express that. This is why the EU approach, applied to agents, tends to produce a compliance artifact that sits beside the system, while the Chinese approach produces one that has to be compiled into it. Whether that was the drafters' intent or a happy accident of translating 分级分类 governance onto a new technology, it is the more architecturally honest cut.

The obligations still scale by sector in the familiar way. Agents in healthcare, transport, media and public safety face filing, compliance testing and product-recall provisions; entertainment and routine office use get self-assessment tooling, information reporting and industry self-discipline. What does not scale away is the authorisation boundary itself.

What to do about it

The obvious action if you operate in China is to read the document properly with counsel. The less obvious and more broadly useful action is to build the artifact it demands regardless of where you ship, because the work is small and it is load-bearing for reasons that have nothing to do with any regulator.

StepWhat it producesWhy it pays outside China too
Enumerate every action your agent can take, including via tool serversA list, usually longer than expectedMost teams discover actions nobody knew were reachable
Classify each into the three tiers by reversibility and blast radiusThe authorisation tableSettles autonomy arguments once, in writing, instead of per incident
Enforce tier 1 and 2 in credentials and a gate, not in the promptAn authorisation check outside the modelThe only defence that survives prompt injection
Record proposal, tier, grant and outcome per actionAn append-only decision logThe evidence any incident review or audit will ask for
Expire standing grants and re-askBounded delegationStops scope creep as the agent gains tools

Be clear-eyed about the limits of the document too. Implementation Opinions are a policy instrument rather than binding law with defined penalties; the enforcement mechanics, the testing standards and the filing procedures are largely delegated to sector regulators and standards bodies that have not all published yet. Anyone telling you exactly what compliance looks like in operational detail is ahead of the available facts. What is safe to conclude is directional: the first national regulator to write down what an agent's autonomy boundary must look like chose per-decision authorisation, and regulators copy each other.

FAQ

Does this apply to a company outside China?

Directly, it applies to agents deployed for users in China, and the practical test is where your users are rather than where your company is. Indirectly it matters everywhere, because it is the first concrete regulatory specification of an agent autonomy boundary and it is likely to shape what other regulators ask for.

Is this a law with penalties?

Implementation Opinions are a policy directive issued jointly by ministries, not a statute passed by the legislature. They set required practice and direct sector regulators to act; the enforceable detail arrives through the sector rules, standards and filing regimes they instruct. Treat it as binding on practice and provisional on penalty.

Is a confirmation dialog enough for tier 2?

Only if the confirmation is what actually unblocks execution. A dialog that the model can decide to skip is UI, not authorisation. The test is whether an action can reach the outside world when no approval record exists — if it can, you have a prompt-level boundary wearing a gate's clothing.

Where does this leave fully autonomous agents?

Unaffected in principle, constrained in practice. Tier 3 explicitly exists, and an agent may act on its own within a delegated scope. What the document rules out is an unbounded delegated scope — the user has to have granted something specific, and the agent has to be unable to exceed it.

We already do human-in-the-loop approvals. Is that the same thing?

It is the same idea and usually a weaker implementation. The common pattern gates on a model-produced confidence signal or on a prompt instruction; the requirement here is that the classification be a property of the action, decided before deployment and enforced independently of the model.

Further reading

On this wiki:

Sources: