Field Notes
As AI tools start asking other AI tools to critique their plans, tests, and output, the humane question is whether the second opinion is defending a real standard or just making uncertainty look managed.
As AI work moves out of chat and into canvases, sites, agents windows, and annotated artifacts, the humane design question is whether people can still grab the work where judgment actually happens.
As AI agents learn to plan before they act, the humane design problem is not making the plan look authoritative. It is making sure somebody can still notice what the plan leaves out.
As people ask AI systems what to do with their jobs, bodies, money, children, and relationships, the humane design problem is not better confidence. It is better doubt.
As AI tools learn from the relationships, files, meetings, messages, and permissions around work, the humane design question is no longer only what context they can see. It is how politely they use it.
As coding agents move into knowledge work, more teams will make small tools for small moments. The hard part is deciding what deserves to keep existing.
As AI agents get identities, permissions, sponsors, and audit trails, the important design question is not only what they can do. It is how they become legible members of the workplace.
As agents take on more execution, the humane workplace question is not whether humans stay in the loop. It is what kind of loop we are asking them to live inside.
As AI moves from answering beside our files to acting inside them, documents, spreadsheets, and decks are becoming rooms where judgment, authority, and evidence have to coexist.
As assistants remember across chats, projects, repositories, and connected tools, the humane design question is no longer whether memory helps. It is where memory should stop.
As AI search becomes a place where answers, interfaces, planning, and transactions happen, the humane design question is whether people can still leave the answer cleanly.
As agents move from chat into managed work environments, the most important interface may be the room we let them work inside.
As AI agents write code, sign commits, and leave traces through our tools, authorship stops being a footer. It becomes part of whether work can be trusted.
As agents begin to run from Markdown files, skills, goals, and workflow rules, the quality of ordinary instructions starts to matter like software architecture.
AI tools may not make everyone a software engineer. Their more interesting promise is that they let more people bring a serious idea close enough to software to be heard.
As AI review moves from comments into suggested fixes and agentic repair, the important question is whether teams still have enough friction to notice what quality actually requires.
As AI work moves from active sessions into phones, schedules, locked computers, and remote agent threads, the humane design problem is no longer only how work begins. It is how it stops.
As software gets redesigned for agents as well as humans, the hidden work is not making everything autonomous. It is keeping systems legible enough to trust.
As AI moves into default models, keyboards, cursors, widgets, and operating systems, the humane design problem is no longer access. It is knowing when help should slow itself down.
The first labor problem may not be that AI eliminates entire professions. It may be that it quietly removes the messy early work where people learn how judgment is made.
As agents move into Slack channels, schedules, and shared directories, the workplace conversation layer is becoming a place where work gets triggered, reviewed, and quietly delegated.
As AI collapses the distance between canvas and code, the scarce work is no longer translation. It is knowing what should survive the loop.
As AI work becomes metered by team, model, feature, and cost, the question is no longer whether people use the tools. It is what the organization teaches them to value.
The important shift is not that AI can summarize faster. It is that research is starting to look like delegated labor instead of assisted search.
The breakthrough is not one more model. It is a shared way for agents to reach tools, data, and systems without every product pretending to be a closed universe.
The frontier is moving away from clever wording and toward the design of what the model can see, retrieve, compress, and ignore.
The model may still be the engine, but the product advantage is moving toward packaged capability: tools, rules, workflows, and embedded know-how.
The browser used to be where we did the work. Increasingly it is becoming the place where we supervise work being done on our behalf.
When text, images, audio, and video can live in the same retrieval space, search stops being a library query and starts becoming a memory system.
The real bottleneck in software was never just typing speed. It was the hidden work of scope, delegation, review, context, and trust.
When multiple frontier agents live inside the same product, model selection stops feeling like loyalty and starts feeling like workflow design.
The tools are getting better at acting, researching, coding, and remembering, while many organizations still think work is a matter of attendance, output, and managerial visibility.
After the fascination with speed, the durable differentiator may be whether a tool helps people produce work that is calmer, cleaner, and more worth trusting.
OKRs can name a destination, but they often miss the conditions that make health, judgment, and quality possible in the first place.
A useful routine generated by a model can still fail on contact with energy levels, mood, weather, and the older operating system known as a body.
The shift was not toward harsher discipline, but toward better systems, fewer identity dramas, and more attention to what my days were already teaching me.
Scarcity, anticipation, pacing, and surprise are not side effects. They are the architecture of how systems become memorable.