Get in touch
← Writing

Essay · 15 October 2026 · Ron Bronson

Towards a Consumer Agent Experience

The person an agent works for deserves a way to see what was done in their name.

Read the Markdown source

In January 2025, Mathias Biilmann, the CEO of Netlify, gave a name to something a lot of developers were already feeling. AI agents had become users of software in their own right, and the products agents found easy to use were going to win. He called it agent experience, or AX.

A year later, in One Year of AX, he organized the field into four areas: access, context, tools and orchestration. The first area asks whether the human must be in the loop.

I started on that question shortly after his first post. On February 2, 2025, I published Service Design for AI: The Disintermediation Problem, about what people lose when an agent stands between them and the services they use. I argued that the industry was optimizing agent experience and needed a matching human experience, or HX, that tracks how agents read intent and where their assumptions put people at risk. I kept writing about it through 2025 and 2026.

Agent experience has a second customer, the person the agent works for. That person needs a way to see what agents are doing for them, what they approved and what actually happened. I call that consumer agent experience. Someone outside the companies building agents can build it now, from specs that already exist.

The person in the transaction

Say an assistant tells you your business license renewal is done. That one word could mean it found the right form, filled it out, sent it, or that the county accepted it. Each of those is a different fact with a different next step.

Today the person who asked has almost no way to tell them apart. They log into the county portal, their email, their bank and the assistant's own history, and stitch the story together themselves.

Now add a calendar helper, a shopping agent, an insurance appeal and a school enrollment form, each from a different company. The person is the only place the whole picture exists, and they keep it in their head.

Waiting for the rails

Standards bodies and governments have started on parts of this. In February 2026, NIST launched an AI Agent Standards Initiative focused on agent identity and authorization. In June, Estonia said it would issue digital identities for AI agents, with limits a person can set, like view only or a spending cap.

Public institutions work on a timeline of years. I spent a long stretch of my career at 18F, across three administrations, and I know why that pace exists. Companies shipping agents release something new every month.

In August, Cloudflare published a plan for an agentic internet where every paid interaction leaves a receipt, and the person's role is described as a wallet and a budget the human set once. In September, Visa, Mastercard and Ant International announced a shared framework for verifying agents. In October, Sierra and Meta published the Personal Agent Protocol, which gives companies a standard way to tell agents what they may do. Its own spec says a company can't confirm the person ever saw the terms they approved.

A person waiting on a public mandate gets no record of their own while all of this ships.

Why no single company builds it

Each of those efforts produces part of the record I'm describing, such as a receipt, a signed approval or an operation status, and each part stays with the company that issued it. Google's Universal Cart follows you across Search, Gemini, YouTube and Gmail, and stays inside Google. Payment networks see payments, and an assistant sees its own conversations.

A record that crosses all of them is valuable to the person and to the market as a whole. It also pays no single vendor to build it, because the value shows up exactly where each vendor's product ends.

So I think this has to come from outside the ecosystem. It can be built from the published specs and attached to whichever rails exist, and the person whose name is on the account owns it.

What I've been writing toward

When I line up what I've published since early 2025, it reads like one argument written out of order. Each post handled one piece, and I couldn't fully say how they fit until I tried to design the thing they describe.

Date Post The piece it added
Feb 2, 2025 Service Design for AI: The Disintermediation Problem What people lose when an agent stands between them and a service, and human experience (HX) as the person's side of AX
Feb 23, 2025 Context-Aware Agents: Designing for Alignment in AX Agents mediate reality, so they need to know why a task exists inside a system, and success needs measures beyond speed and completion
Apr 9, 2025 Not a Roadmap When the interface fades, so does the path back, and the question of who decides whose outcomes an agent optimizes for
Oct 6, 2025 Context Is Not Just a Map, It's the Terrain People handing sensitive decisions to agents, then working across other agents to fix the mistakes
Jan 15, 2026 Who Holds the Pen? Judgment routers that route decisions, four tests for when a person signs off (uncertainty, stakes, authority, novelty), and decision receipts that tie human authority to agent action
Feb 1, 2026 On Anti-Delegation Our knowledge sitting in storage units we don't own, which can be abandoned and sold off
May 3, 2026 Administrative Burden As a Service Judging the work by whether the person who needs help actually gets it, on time
Jun 11, 2026 An after-action review skill A structured debrief of what an agent did and what a person authorized
Jun 13, 2026 Skill Issue Is Not a Product Strategy People absorb the cost when these systems fail, and the retry is part of the business model
Jun 17, 2026 When AI Agents Meet Insurance Agents An agent used to mean a person authorized to act for you and answerable for it, with trust running both ways
Jun 29, 2026 closedtab A six part record of agent work: intent, action, judgment, deviation, consequence, change
Aug 9, 2026 Agent Experience Needs Failure Affordances A way for an agent to stop and raise its hand when something is off
Oct 5, 2026 Dark Patterns Won New technology grafting itself onto existing rails without paying for their upkeep

Read in order, the questions go from what a person loses when an agent stands in, to whose outcomes the agent serves, to who decides, to who keeps the record, to how an agent stops, to who the record belongs to. The code is public too: closedtab, judgment-routing and the after-action review skill.

What the record has to do

A person connects the helpers that act for them and the places those helpers work, sets limits the way they'd set a spending cap on a card, approves the actions that need their yes, and keeps a history of everything that happened.

The part I care about most is how that history treats evidence. It keeps four things separate:

  1. What the helper says it did
  2. What went out and what came back
  3. What the company confirmed
  4. Whether the result actually shows up

When those disagree, the disagreement lands in front of the person as something to resolve. Without that, a person finds out weeks later, from a bill or a missed deadline.

A record like this can attach to whatever rails companies publish, whether that's the Personal Agent Protocol, Google's payment mandates, MCP or a plain API, and it keeps working whichever of them wins.

Why write it down now

Agent experience so far has focused on making software easy for agents to use. Consumer agent experience adds what the person gets back: a view of their helpers, the limits they set, and a record they can carry from one helper to the next. People can have that view today, while laws and rails catch up.

I'm putting this on the record with dates, the same way a good record of delegated work would. If people come to expect that record, I want it clear where the idea came from and how it got built.

Sources