Practitioner's Corner
Practitioner's Corner
The Next Translation Layer
The IRS's Individual Master File has been running since 1961. It was supposed to be replaced decades ago. Instead, layer after layer of new technology grew around it—databases, portals, APIs—each one promising to make the last one obsolete. None did. And at some point, the logic that kept things working migrated out of the code and into the people operating it. The workarounds became the process.
Now AI agents are arriving as the next modernization layer, and the promise sounds familiar: this time, the old dependencies finally break. But agents have to interact with the same systems, the same fragile sequences, the same fields that reject data until someone knows the trick.
The Next Translation Layer
The IRS's Individual Master File has been running since 1961. It was supposed to be replaced decades ago. Instead, layer after layer of new technology grew around it—databases, portals, APIs—each one promising to make the last one obsolete. None did. And at some point, the logic that kept things working migrated out of the code and into the people operating it. The workarounds became the process.
Now AI agents are arriving as the next modernization layer, and the promise sounds familiar: this time, the old dependencies finally break. But agents have to interact with the same systems, the same fragile sequences, the same fields that reject data until someone knows the trick.
Building Browser Agents for a Web That Fights Back
A government portal in Delaware goes offline at night. A dropdown is actually a textbox. A checkbox arrives pre-checked, hoping nobody notices. The web is a collection of pages to read and forms to fill. It is also, for anything trying to automate those forms, an adversary: elements misidentified, structures obfuscated, layouts rearranging between visits. Most browser agents get built as though the environment cooperates.
Suchintan Singh and Shuchang Zheng failed at two startups before they noticed what kept breaking was every automation they wrapped around their products. That observation led somewhere uncomfortable.
Building Browser Agents for a Web That Fights Back
A government portal in Delaware goes offline at night. A dropdown is actually a textbox. A checkbox arrives pre-checked, hoping nobody notices. The web is a collection of pages to read and forms to fill. It is also, for anything trying to automate those forms, an adversary: elements misidentified, structures obfuscated, layouts rearranging between visits. Most browser agents get built as though the environment cooperates.
Suchintan Singh and Shuchang Zheng failed at two startups before they noticed what kept breaking was every automation they wrapped around their products. That observation led somewhere uncomfortable.
The Broken Simulation
Every enterprise AI pitch starts the same way: modernize the stack, add APIs, escape the legacy mess. Amazon's AGI Lab is running the opposite play. Its researchers build reinforcement learning gyms that faithfully reproduce decades-old software, quirks included. Modal windows appearing late. Fields rejecting input until some other value saves first. Forms silently resetting midflow.
The lab treats these behaviors as the real semantics of the system. Train an agent inside that friction long enough, and you get a synthetic API over infrastructure nobody ever designed to be programmatic.
Further Reading
