Seeing Is Believing

In my writings about faith, I have explored how hard it is for humans to intuitively believe in what we do not see. Most of us will sit in a chair without inspecting it, but when it comes to our money...the story is way different. We're a lot more risk-averse, and that difference sets up a bank story for one of my favorite buildings in NYC.

The Vault in the Window

For most of the last century, banks earned trust through mass. Granite columns, heavy doors, and a vault hidden in the basement told depositors their money was safe without ever showing them how. In movies and cartoons, we see robbers needing drills, lasers, and dynamite to get through many layers of defense. Fort Knox was meant to make you feel your money was protected in an impenetrable bastion.

In 1954, Manufacturers Trust Company opened a branch at 510 Fifth Avenue and decided to go against the grain by doing something daring, showing clients the exact thing protecting their money. Gordon Bunshaft of Skidmore, Owings & Merrill wrapped the building in floor-to-ceiling glass and moved the vault out of the basement to sit right behind the window, where anyone walking down Fifth Avenue could see its circular steel door—7 feet across, 30 tons, and designed by Henry Dreyfuss with the Mosler Safe Company. On its first day, 15,000 curious people stopped by the branch, and Manufacturers Trust's deposits tripled within nine months of the opening. Is this why PMs at Apple thought glass mode would be such a killer feature?

Small fun fact: Manufacturers Trust later became Manufacturers Hanover, which merged with Chemical Bank, which merged into Chase, so the building was once owned by the bank I work for today.

I love thinking about this story because it maps directly to what I have experienced in AI recently. I trust output from agents more when I see it working. Within a terminal session, I step away and Claude Code says the task is done. The summary reads confidently, and naturally I go test the exact logic instead of believing the model output. For the majority of cases, things are good, but in some cases even that isn't enough. The trouble with a probabilistic system reporting its own success is that the report and the work come from the same place. I noticed particularly during Opus 4.8 sessions earlier this year that the model would sometimes just hardcode the outcome I wanted to see! It'd initially pass my sniff test, but any other input would show that the emperor had no clothes.

The Emperor in the Side Panel

Anthropic shipped a pretty cool feature a few months ago...akin to putting the vault in the window. In the desktop app and on the web, Claude Code can open a real browser in a side panel next to the conversation, and I can watch that pane while it works. In April 2026, the pane was limited to showing what the agent and harness were implementing. For example, when it changes a web app, it can start the dev server, load the page, click through the change, read the console for errors, and take a screenshot as proof. If something looks off, I can take over right there.

I am a naturally stubborn person, and it took me 3 months to leave my beloved terminal interface behind (partly because I was also running Codex in parallel). As a self-confessed lousy engineer, I found that seeing the changes live, as opposed to reading the chain of thought and code diffs, changed everything for me. I can tell pretty reliably when a page looks right, or at least whether the DTOs coming from the back end are mapping correctly to the front end.

The agent still tells me what it did, and now the evidence sits right next to the claim, so the agent is no longer the only one who saw the outcome. Around the same time, Anthropic shipped new functionality that lets the browser pane show the agent browsing. That is, I can see what websites Claude is reading for context, which page of technical documentation is feeding into the implementation, and how that is flowing into the shape of the software that's being assembled in the session. Once I could see how the work is done, my confidence in Claude's output and my ability to steer midway 10x'ed overnight.

At work, most of the conversation around agents across different teams is about making them more capable, and much less is about making their work easier to check. I believe that focus is misaligned for high-stakes agents in the treasury space, since capability without verification actually moves the burden onto the user, who has to either blindly trust that the money moved correctly or reconcile through reporting after the money has left.

A Reconciliation Layer for Agents

Which brings me to the outsized impact of what my team is building. We are bringing personalized and intelligent agents surfaced via a chat interface to Fortune 100 treasury teams. A wrong answer about a payment can mean a supplier goes unpaid, a cash position gets misread, or a treasurer makes a decision on bad information.

In Trust Is the Rail, I wrote that payment systems flow freely when trust is concentrated and are forced to reconcile when trust is distributed. An agent sitting between a client and their money is a new trust boundary, so we are designing for reconciliation from the start.

The pattern is simple to describe but tedious to build. Every time our agent acts or answers, it does so through tool calls against multiple API layers that wrap systems of record. There are entire control planes enforced around enterprise-grade entitlement and provisioning layers, and when the work is finally done, the agent navigates the user to the screen in J.P. Morgan Access where the outcome actually lives. If a user asks the agent to find a payment, they land on that payment's detail page. If they ask it to amend an instruction, they land on the amended instruction. In other words, the chat response points the user to the answer instead of asking them to accept it as the answer.

What I like most about this approach is that we aren't asking clients to trust anything new. They already work in Access every day, and they already believe what those pages tell them. The agent borrows that existing trust instead of asking for fresh trust in a chat bubble, and if the agent's summary and the Access page ever disagree, the user can see it right away.

So what's the tradeoff? For now, our experience isn't Pareto dominant. Navigating the user to a confirmation screen adds a step, and it makes the experience feel less magical than a confident one-line answer. On the engineering side, every tool needs a mapped destination, every destination needs a reliable deep link with the right entitlements, and some outcomes require state or parameters to be passed in the URL too. It would be much faster to ship an agent that just tells users what happened, but we would be asking them to believe us without showing them.

When Users Stop Looking

The hope is that, after enough times seeing the agent's claim match the page, users won't feel the need to look anymore. Trust gets built the way it always has, through repeated agreement between what someone says and what you can verify, until verifying starts to feel unnecessary. Showing the work every time is how we get there, and I don't think of it as a permanent feature of the product.

I suspect one useful signal will be how often users actually open the confirmation screen over time, and whether that drops as they get more comfortable. If it falls, I would read that as the reconciliation layer doing its job.

The Door Without a Vault

Back on Fifth Avenue, if you walk past the building you can't actually see the vault today. At some point in time, partitions went in, and it looks like the building was converted to retail. The vault's walls were removed but the door is still there today.

I'd like to think that both halves of the story will apply to the agents we built. Almost nobody walks into a bank to look at a safe anymore. We check a balance on our phones and believe it, because decades of that number being right removed the need. That is the future I hope for, where clients stop needing to watch the work in between.

For now, we have to put the vault in the window. I honestly don't know how long it will take before our clients feel comfortable looking away, and I expect it will happen one workflow at a time rather than all at once. All the fun of building for trust at a G-SIB. If any of this seems like an interesting challenge to you, please reach out to me. Our team is always on the lookout for exceptional first-principles thinkers.