It keeps getting better

Two things have been improving at the same time.

First, the obvious one: artificial intelligence has become much more capable. It can reason across a codebase, work with images and documents, inspect interfaces, research unfamiliar territory, and carry a task much further than it could a year ago.

Then there is the less obvious improvement. The system around the model has grown up. Codex moved from an editor panel and terminal into a desktop environment with files, images, a live browser, isolated worktrees, plans, persistent goals, memory, and agents that can take bounded pieces of work.

Code.Files and images.The live browser.Worktrees and goals.Memory and agents.

The snapshot above catches that change in motion. Live browser context moved from an occasional attachment to something present in most recent messages. Goals and bounded agents appeared as a normal part of the system almost at once.

Some of that jump is new functionality becoming available, so it is not a clean productivity comparison. That is exactly the point. A new capability appears, I start using it, and a few weeks later it has changed the shape of the work around it.

In September, 155 of 168 classifiable threads were Selfso. In January it was 367 of 391. By the first 13 days of July, Superhuge led with 63 threads, Selfso had 56, and horse search, course design, Shopify, perpetual futures (perps), and other projects were active around them.

The range matters more than the count. On Selfso, Codex helped carry native notification work through backend rules, Swift, Kubernetes, TestFlight, and real Apple Push Notification service behaviour on a physical phone. The code was only one part of getting to true.

On Superhuge, I use it as a visual and editorial partner. We work on shaders, motion, layouts, copy, project stories, and this note. The live browser stays open because the last bit is often feeling, not a test result.

For a horse search, it helped turn scattered public information into a sourcing system with 23 sheets and 919 rows. In perps sessions it refreshes the actual market and account, and sometimes the best result is no trade.

Same environment. Different evidence, different risk, different definition of done. The useful skill is not asking AI to code. It is learning how to build the right loop around whatever kind of work is in front of me.

My current loop usually starts with the real thing: the route, device, workbook, theme, account, or production state. We inspect before editing when the problem is unclear. We make decisions, record the ones that need to survive, implement a bounded slice, run checks, and return to reality.

Even that is new. Plans came first. Then isolated worktrees made parallel code lanes safer. Persistent goals turned a task into a small operating contract with an objective, non-goals, and a finish line. Agents now let one task investigate while another implements or verifies.

I am not throwing work at a cloud of unsupervised bots. The better the system gets, the more specific I can be about ownership, evidence, stopping points, and handoffs.

Each improvement removes a bit of re-entry and a bit of translation. The work can move further without losing its history. That makes bigger loops possible for one person.

Six months from now this description will probably feel dated. Excellent.