Compliance Is Architecture: The Five Questions the Laptop Answered for Free
The word "compliance" appears zero times in Anthropic's managed-agents post and zero times in Brex's autonomous-agents post. Of the three reports this handbook is built on, only Intercom covers it, and they covered it because 93% of their pull requests are agent-driven and their auditors needed evidence. That silence is not an oversight by two otherwise rigorous engineering teams. The laptop model made compliance the human user's problem by construction, so there was nothing left for a vendor to address. When the agent runs on your machine, against your checkout, committing with your GitHub token, "who's responsible?" is trivially "the human at the keyboard." The vendor never has to think about it.
Once agents leave the laptop, compliance is a set of properties the architecture has, not a certificate you bolt on.
The demand that forces it is documented. Eight in ten organizations are already deploying or planning to deploy agentic AI (Vanta), and the goal teams set is not "a few PRs": Intercom now ships over 93% of its PRs agent-driven (Intercom). A team that wants half its code shipped by AI and wants to stay audit-ready is asking for exactly the thing the laptop gave away for free.
For one engineer, the laptop is a complete answer
The dodge works because the laptop model is genuinely sound at the scale it was designed for. For a single engineer on a managed device, every compliance question collapses onto one person, and that person is already covered by controls you have. Identity? Your token signed the commit. Access? Whatever your SSO already provisioned, already reviewed in your last access audit. Residency? A corporate device inside the audited perimeter. Approval? You read the diff before you pushed. Evidence? The git log, under your name.
Five questions, one answer: you. The vendor inherits a decade of endpoint management, identity governance, and change-management process without shipping any of it. It is a free ride, and for an individual it is the right call. See why the laptop model breaks for the operational half of this argument.
Five questions reopen the moment agents leave laptops
The collapse depended on one physical fact: a human was present at every change, on a device bound to their identity. Move agents into shared cloud environments, which you must do to get predictability, scale, and parallelism, and that fact is gone. Every collapsed question reopens, and each one now needs a real answer instead of "you."
| Question | Laptop answer | What the cloud demands |
|---|---|---|
| Whose identity is on a commit an AI wrote? | Your token | A distinct AI identity, bound to the initiating human |
| What did the agent have access to, and was it the minimum? | Your ambient access | Per-task scoping, credentials kept out of the sandbox |
| Where did the customer's code physically run? | Your corporate device | A placement policy you can state and prove |
| Who approved the change, and can they be the writer? | You, implicitly | An enforced gate where author and approver differ |
| Can you prove any of this, after the fact, for every task? | The git log | An external, queryable record of every step |
None of these questions is new; what is new is that no single human can stand in as the answer to all five. Brex states the attribution requirement plainly: agent work "needs to be attributed back to an actual human and not just a bot," and agents "should only have access to the information they need." Those are two of the five questions, named by a team that hit them in production.
"You're the responsible human" doesn't scale to thousands of changes
Half your code shipped by AI means answering all five questions for thousands of changes, without a human babysitting each one. Intercom pushed 497 fully autonomous PRs to production in the first four weeks of broader rollout, with Claude writing the code and an AI approval system reviewing and shipping it. No per-change human attestation survives that volume. It becomes rubber-stamping with extra steps, the exact failure approval fatigue describes.
The industry's current posture confirms the gap. Gravitee's 2026 survey of 900+ executives and practitioners found only 21.9% of respondents treat AI agents as independent, identity-bearing entities in their security model. Vanta reports that 70% of companies have AI tools accessing their environment without going through procurement, and fewer than 2% of unmanaged vendors ever receive a security review. IBM's 2025 Cost of a Data Breach report found 97% of organizations reporting an AI-related security incident lacked proper AI access controls.
One fence worth marking: Intercom's post is the only published first-party account of taking agent-driven change management through real audit frameworks. The field is otherwise silent on what an auditor actually accepted.
Compliance is architecture, not policy
A system has to carry the responsibility a human can no longer carry per-change, which makes compliance a set of properties the architecture has rather than a certificate you bolt on. Intercom demonstrates the shape. Every AI-approved PR is "labelled, logged, and queryable," they engaged their auditors (Schellman) before scaling, and they confirmed the evidence meets SOC 2, HIPAA, ISO 27001, ISO 42001, and AIUC-1. Their conclusion is the thesis of this whole section: "When you build for safety, compliance follows."
The five reopened questions resolve into seven properties the environment must have, and a system built for those properties produces the auditor's evidence as a byproduct of normal operation rather than as a separate scramble before each audit.