Reading draft agent legislation for architectural requirements
Legislative text changes between draft and enactment. What tends not to change is the underlying question a bill is trying to answer — and for agent legislation that question is consistently about accountability for automated acts.
What should you build from draft agent legislation?
The four demands that recur across every draft: identification, human accountability, records and revocation. Thresholds and definitions change repeatedly during drafting; the problem being responded to does not. Three of the four are cheap, and one cannot be retrofitted.
- Draft agent bills converge on identification, human accountability, records and revocation — across jurisdictions and drafting styles.
- Those four are architectural. Retrofitting them into a deployed agent estate is substantially harder than designing for them.
- Building to the convergent requirements is defensible engineering even if no specific bill passes, because the questions they encode get asked by counterparties and insurers regardless.
Part of Agent tool-call gating
Reading bills for the question, not the text
Specific thresholds, definitions and penalties change repeatedly during drafting. The problem a bill is responding to changes rarely, and it is the problem that determines what you have to build.
For agent legislation the problem is consistent: automated systems take consequential actions and nobody is clearly accountable for them.
The four convergent requirements
| Requirement | What a bill typically asks | What it means architecturally |
|---|---|---|
| Identification | Agents must be identifiable as such to counterparties | A stable agent identity, distinguishable from a human's |
| Human accountability | A named person or entity is responsible for the agent's acts | A record connecting the agent to a principal, per action or per scope |
| Records | Sufficient records to reconstruct an automated decision | Execution records independent of the agent's own narration |
| Revocation | Authority can be withdrawn and the withdrawal is effective | Enforced revocation at the point of effect, not advisory |
These appear in draft text from several jurisdictions with different legal traditions and different drafting conventions. The convergence is a signal that they are responses to the structure of the problem rather than one legislature's preference.
Which are cheap and which are not
- Identification is cheap. Issue distinct identities, label agent traffic, do not let agents share human accounts. Mostly hygiene you should want anyway.
- Records are moderate. Capturing what executed, in canonical form, at the point of effect. Storage is trivial; the work is instrumenting the right points.
- Revocation is moderate. Enforced at the effect rather than the gateway, failing closed. A day's work in a well-structured service, a month in a sprawling one.
- Human accountability is expensive to retrofit. It requires a record connecting a named human to an action or a scope, produced at the time and verifiable afterwards. Systems that did not capture it cannot reconstruct it.
The asymmetry in item four is the whole argument for acting before enactment. The other three can be added to a running system. The fourth cannot be backdated, because the human was not asked.
What "human accountability" does not mean
It does not mean a human approves every action. No draft requires that, and any that did would be unworkable.
It means the chain from an action to a responsible human is traceable. For most actions that runs through a scope grant — a human authorised this class of action within these bounds. For a small set of high-consequence actions it runs through a per-action approval.
Conflating the two produces either paralysis or a control that does not hold.
Building without betting on a bill
The honest framing: do not build to a draft. Build to the convergent requirements, which are demanded by parties other than legislatures.
- Counterparties ask whether your agent had authority before they perform
- Insurers ask what controls governed automated actions when pricing
- Auditors ask how automated decisions are evidenced under existing frameworks
- Your own incident response asks what actually executed
Each of those arrives regardless of whether any particular bill becomes law, and each is answered by the same four capabilities.
A reasonable sequence
- Inventory agents and the consequential actions they can take. This is usually the step that surprises people.
- Give agents distinct identities. Stop credential sharing with human accounts.
- Instrument execution records at the point of effect for the consequential set.
- Implement enforced revocation with a short kill time.
- Add per-action human approval on the irreversible subset, bound to the rendered effect.
- Add scope grants for everything else.
Steps one through four are defensible under any regulatory outcome. Steps five and six are where the effort concentrates, and where a system built without them will struggle later.
This describes legislative tendencies at a general level and is not legal advice. Specific obligations depend on jurisdiction, sector and enacted text.
Why build without betting on a bill
| Party | When |
|---|---|
| Counterparties | Before performing on an agent commitment |
| Insurers | When pricing, and at claim |
| Auditors | Under existing frameworks, now |
| Your own incident response | At 3am |
Objections and honest limits
“Human accountability means approving everything.” No draft requires that and it would be unworkable. It means the chain from an action to a responsible human is traceable — usually through a scope grant, per-action only on irreversible effects.
“We should wait for enacted text.” Identification, records and revocation can be added to a running system. Accountability cannot be backdated, which is the asymmetry that decides the sequencing.
This describes legislative tendencies at a general level and is not legal advice. Obligations depend on jurisdiction, sector and enacted text.
A sequence that survives any outcome
- Inventory agents and their consequential actions. The step most often skipped.
- Give agents distinct identities. Stop credential sharing with human accounts.
- Instrument execution records at the point of effect. For the consequential set.
- Implement enforced revocation. Short kill time, fails closed.
- Add per-action approval on the irreversible subset. Where effort concentrates.
- Add scope grants for everything else. So the chain terminates at a person.
Terms used here
- Convergent requirement
- A demand appearing across drafts from different jurisdictions, indicating it responds to the problem's structure.
- Scope grant
- A signed delegation covering a class of actions with bounds and an expiry.
- Retrofit cost
- How expensive a capability is to add later — the variable that should drive sequencing.
Frequently asked questions
Should we build to a bill that has not passed? No. Build to the requirements that recur across drafts, because counterparties, insurers and auditors ask the same questions regardless of enactment.
Does human accountability mean approving every action? No. It means the chain from an action to a responsible human is traceable — usually through a scope grant, and through per-action approval only on irreversible effects.
Which requirement is hardest to retrofit? Human accountability. Identification, records and revocation can be added to a running system; a record of what a human authorised cannot be created after the fact.
Is this legal advice? No. It describes tendencies in draft legislation at a general level. Specific obligations depend on jurisdiction, sector and enacted text.
Should we build to a draft bill? No. Build to the requirements that recur across drafts, because counterparties, insurers and auditors ask the same questions regardless.
Does accountability mean approving every action? No. It means the chain from an action to a responsible human is traceable, usually via a scope grant.
Where this fits in Manav
Manav covers three of the four directly: distinct agent identities under a signed human delegation, execution records at the point of effect, and revocation enforced where the effect happens.
Sources and further reading
- Published draft legislation on autonomous software systems and AI agents.
- Regulation (EU) 2024/1689 (AI Act) — consolidated text
- Sector regulator guidance on accountability for automated decisions.
- Comparative analyses of AI governance requirements across jurisdictions.