Notes from the Intersection No. 003
Carmen Onchain · 05 June 2026 · 8 min read
When AI Agents Make Decisions, Someone Has to Answer for Them
Picture a Friday morning in a mid-sized bank somewhere in Asia. A treasury team logs in and finds that an AI agent rebalanced a multi-million dollar position overnight, in line with its mandate, in line with the data, and outside the loss limits the team thought it had set. The agent did not malfunction. It did exactly what the deployer asked it to do. The deployer is now reading the policy document for the first time. The counterparty is on the line. The regulator is forty minutes away.
This scenario is no longer hypothetical. It is the operating model that enterprise software is quietly sliding into across finance, hiring, healthcare, and logistics. The week of 2 June 2026 made the slide visible in a way you cannot unsee, because three different jurisdictions and three different vendors all moved at once, in opposite directions, and the gap between them is now the most interesting place to build.
The question is not whether AI agents will run consequential decisions. They already do. The question is who answers when one of those decisions goes wrong, and what they hand the person across the table to prove what happened.
If a regulator landed on your stack tomorrow and asked you to prove which agent made which decision on whose authority, what would you hand them?
Deep Dive
Decision Infrastructure Is Here. The Rulebook Is Not.
What is actually happening here
Singapore wrote the first real liability map for AI agents this week. The framework, published by the Infocomm Media Development Authority and covered by Complete AI Training on 1 June 2026, breaks the value chain into clearly labeled roles. The developer who builds the agent has one set of obligations. The deployer who puts it into production has another. The auditor has a third. The end user has the smallest and most specific. Each role gets a fence around what it owes the others.
The same week, the United States went in a different direction. President Trump signed an executive order asking federal agencies to review the security implications of high-risk AI models. The National Law Review, in its coverage on 5 June 2026, called it a directional instrument rather than an enforcement one. Connecticut passed a separate law requiring employers to disclose AI use in hiring decisions. Congress introduced a bill to limit military use of AI. None of these instruments cite each other. None of them define their key terms the same way.
Meanwhile, the vendors moved. Nvidia spent the week stacking agentic AI infrastructure across three product lines, as CIO covered on 2 June 2026. Asana and Palo Alto Networks each made acquisitions to bolt execution and security layers onto their enterprise stacks. Analytics Insight described the underlying shift bluntly on 5 June 2026: AI is becoming decision infrastructure rather than software you launch.
Why this matters beyond the money
It is tempting to read this as a regulation story, and to file it next to every other piece you have read about the EU AI Act or the California patchwork. That misses the shape of the moment. The Singapore framework matters because it does something the other instruments are still avoiding: it accepts that an AI agent is a thing that acts on someone's behalf, and that the law is going to need to follow the action back to a person.
That single move opens a long list of practical questions. If an AI agent declines a job applicant, which party owes that applicant an explanation? If an agent moves money in a way the deployer did not anticipate, can the deployer point to the developer's training data and shift the blame? If an autonomous system harms a third party who never agreed to interact with it, who carries the duty of care? The traditional answer in software is "the company that shipped it." The AI agent answer, the one Singapore is now writing into law, is closer to "it depends, and you should know in advance who is on the hook."
The deeper point, the one most coverage misses, is that liability without a verifiable record is unenforceable. A framework that names roles but cannot show what each role did is just a vocabulary. The record is the part regulators are leaving to the market to build. That is the opening.
What builders and operators should be watching
Three things, in order of urgency. Watch which jurisdictions are writing role-based liability frameworks rather than capability bans. The role-based ones (Singapore, parts of the EU AI Act, draft language in two US states) are the ones builders can design against. The capability bans are noise that will be litigated for years before anyone agrees what they cover.
Watch the audit substrate. The phrase "decision log" is going to do a lot of work in the next eighteen months. The companies that figure out how to produce a tamper-evident, time-stamped, signed record of agent decisions, and to do it without breaking commercial confidentiality, will own the compliance layer of decision infrastructure. This is where blockchain rails stop being a culture-war topic and start being a procurement requirement. A verifiable log does not need to be a public chain. It needs to be a place where you cannot quietly rewrite history.
Watch the vendors. Nvidia, Asana, and Palo Alto are not buying compliance startups yet. They are buying execution and security. The acquisition behavior on the compliance side will tell you when the market believes the rules are about to bite. Blockchain and AI are converging on this question, even when nobody on either side is using the word.
This is the most consequential AI policy document of the week, and almost nobody outside of policy circles is reading it. Singapore did not write a capability ban or a moratorium. It wrote a role map. That is the framework other countries will copy or react to. If you ship anything that touches an AI agent in the next eighteen months, this is the document that decides what your audit obligations look like.
Complete AI Training · 01 June 2026
When the company that already owns the compute layer starts shipping the orchestration layer, the term "AI agent" stops being a product category and starts being plumbing nobody can opt out of.
CIO · 02 June 2026
The decentralized AI movement spent two years arguing about ideology while the centralized stack shipped product. A measured gap is more honest than a slogan, and far easier to close.
Crypto Reporter · 01 June 2026
"Silently" is doing the heavy lifting in that headline, and it is exactly the kind of consent failure that role-based liability frameworks are now being written to catch.
Headtopics · 05 June 2026
Workflow tools designed for agents instead of humans is the operator signal that AI agents are no longer the user, they are the customer.
FinancialContent · 04 June 2026
The story this week is not that AI agents are coming for your job. It is that AI agents are quietly becoming the layer underneath every consequential decision in a regulated industry, and the people who will determine whether that goes well are the ones building the boring compliance plumbing right now.
I keep coming back to a simple thought. The rules will arrive. Most builders will be caught flat-footed. A small group is using this window to build the audit substrate, the role-based logs, the verifiable trail that turns "AI made a decision" from a liability into a documented choice somebody owns. That work is happening. It is winnable. And it is the most useful thing a founder at the blockchain and AI intersection can be doing right now.
What is an AI agent in plain language?
An AI agent is a software system that takes actions on someone's behalf based on goals it is given, without checking back for approval on every step. It differs from a chatbot in that it does things in the world, like sending money, scheduling meetings, or rejecting job applicants, rather than just producing text.
Who is legally responsible when an AI agent causes harm?
As of June 2026, the answer depends on jurisdiction and is in active development. Singapore published a framework on 1 June 2026 that splits responsibility between the developer, the deployer, the auditor, and the end user, with each role assigned specific obligations. The United States has no equivalent federal rule, and individual states are writing their own narrow laws.
Why does blockchain matter for AI agent accountability?
Liability frameworks need a verifiable record of which agent did what on whose authority. Blockchain rails provide a tamper-evident, time-stamped log that an auditor or aggrieved user can read after the fact, without depending on the goodwill of the company being audited. This is the trust layer AI governance has been missing.
What should a founder building with AI agents do this quarter?
Write down every decision your agent can take without a human in the loop and label each one reversible or not. Pick an audit substrate before you scale, something you cannot quietly rewrite. These two steps will put you ahead of most teams when role-based liability rules arrive in your jurisdiction.