CURRENT
A journal for living in the agentic age
Across four decades, every generation declared screen-level automation temporary. Every time, the workaround became infrastructure.
How a browser testing standard's spec language quietly became the operational vocabulary for AI agents on the live web.
Browser automation's twenty-two-year arc traces a single assumption, built for testing known answers, now carrying agents that have none.
RPA's governance lessons were hard-won and well-documented. The teams deploying AI agents today are largely building without that institutional memory.
Two decades of browser automation taught engineers to distrust the gap between command and outcome. Web agents inherited the gap and every failure mode it produces. The skepticism didn't transfer.
Traces how ambient authority created the same vulnerability across four decades, and why LLM agents make the inherited fix harder.
Feedback loops close only when chokepoints, escalating consequences, and reduced compliance costs converge to reach the insulated party.
Why coordination fixes keep winning over structural ones: the frame that prevails is held by people insulated from the failure.
Each workaround built by the people suffering breakage insulates the people causing it, compounding the distance between failure and fix.
How cookies and session state became a de facto identity layer built around human presence, now occupied by agents the architecture never anticipated.
How OAuth's edge-case provision for machine clients quietly became the dominant pattern, collapsing the identity separation the protocol was designed to preserve.
Each era's identity mechanism assumed a simpler caller than the last. AI agents are the first caller that reverses that assumption.
Every generation of CAPTCHA encoded a confident assumption about the human-machine boundary. Every assumption eventually broke. The pattern continues.
A decade of coping with flaky browser tests built infrastructure for environmental chaos. Agents reintroduce non-determinism where those tools can't reach.
Three generations of automation, from mainframe screen scraping to RPA to AI browser agents, each broke the same way for the same structural reason.
Agentic AI is rediscovering RPA's well-documented failure modes without consulting a decade of detailed incident reports.
Integration protocols from CORBA to MCP keep deferring governance, each generation inheriting the unresolved problems of the last.
Technical knowledge encoded in rejected vocabulary becomes unreachable, making rediscovery structurally inevitable when new communities rebuild what BPEL already formalized.
When cookies became blockable, the industry turned browser rendering quirks into identity markers, building on signals that were never stable.
How a 1994 shopping cart fix became so load-bearing that a trillion-dollar company spent six years failing to remove it.
How thirty years of browser impersonation built a self-reported identity layer that AI crawlers now follow to its logical conclusion.
Every proposed successor to robots.txt inherits its core assumption—and the conflict that assumption cannot see.
A 1994 protocol governs web access by reading labels. The new AI agents aren't wearing any.
Robots.txt worked because crawlers returned what they took. AI broke that bargain, and no protocol fixes misaligned incentives.
On-device AI promises to redistribute intelligence from the cloud. The economics suggest both layers expand simultaneously.
Computing's center never actually shifted. Mainframes sank beneath visibility and became the floor everything runs on.
Computing power once moved outward to individuals. AI's cloud architecture is quietly pushing it back. The cycle may be turning again.
Every swing of the rendering pendulum deposits frameworks that never leave. jQuery, AngularJS, Webpack: the weight only accumulates.
Browser engine diversity only decreases. The mechanism that killed Presto and EdgeHTML is still running, and accelerating.
Every era's cheapest rendering location gets narrated as permanent architectural wisdom, until the cost structure shifts again.
BGP hijacks aren't exploits. They're a trust-based protocol working exactly as designed, in a world it never anticipated.
The web's security upgrade required a rare coincidence of concentrated power. Internet routing has no equivalent bottleneck.
A 1989 napkin sketch still routes the global internet, and every patch makes replacing it harder.
From Victorian telegraph wire to fiber optic glass, the world's data still squeezes through the same contested straits.
Nineteenth-century land grants quietly determined where America's internet runs today. Legal permissions outlast every technology they carry.
Telegraph tubes from 1930 now carry fiber optic cable because in communication networks, signals change but geometry persists.
OAuth's authorization flow assumes someone is there to click "Authorize." Agents broke that assumption.
Web sessions were built around humans who walk away. Agents never leave the chair.
Every CAPTCHA you solved trained the machine that made you obsolete at solving CAPTCHAs.
A well-understood timestamp overflow approaches in twelve years, yet billions of devices remain structurally beyond repair.
A single 1970s representation choice created fifty years of security vulnerabilities that awareness alone could never fix.
How a 1990s browser rendering disagreement became a formally maintained specification, revealing backward compatibility as an escalating architectural commitment.
Cloudflare turned the web's governance vacuum into a billing product, and quietly set the default for AI access.
A new licensing format echoes a thirty-year-old convention, betting cooperative norms still work when defection pays better.
A 1994 text file governed web crawling for three decades, until ignoring it became more profitable than reading it.
AI tools promise to finally crack COBOL migration. The pattern they're repeating suggests otherwise.
Y2K was supposed to kill COBOL. The largest software rescue in history proved preservation always wins.
COBOL's fixed-point precision solved financial computing so exactly that sixty-seven years of infrastructure crystallized around the solution.
Your browser choice quietly determines which privacy architecture governs your web, and the split isn't closing.
A 1994 session-state mechanism now sits beneath a billion-dollar consent infrastructure that largely fails at compliance.
A 1994 fix for shopping carts became adtech's invisible backbone, and the people who could remove it never will.
Why IPv6 struggles after thirty years while HTTP/2 succeeded in five—successful workarounds kill protocol transitions faster than technical problems.
How HTTPS went from optional to mandatory in a decade after sitting ready for twenty years—economics and embarrassment, not technology.
Protocol workarounds from the 1990s became permanent infrastructure, creating today's commodity market where IPv4 addresses trade at $35-52 each.
