How Has the Way We Work With AI Changed Over Three Years?
In Anthropic's future-of-work conversation (The future of work with @Claude), which introduces a new feature called Claude Tag, engineers sum up the last three years in three stages. At first, a person decided on a line of code and AI helped with autocomplete. Then one person ran several Claudes at once, handing off work feature by feature. Now AI drives the work directly, and it isn't one person but an entire team taking part in the session.
The key is that the structure flipped: from a human typing one line at a time inside the loop, to AI handling the whole task while the person only sets direction. Over the same span, the time models can work autonomously has grown sharply -- the latest evaluations put it at up to 16 hours in a single stretch.
From Reactive to Resident: What Actually Changes?
Until now AI was reactive: you opened it, asked, took the answer, and pasted it to your team. In the resident model, AI sits inside a collaboration channel, jumps in when needed, and holds onto work that takes days or weeks, following up on its own. It can even finish a 16-hour task and then schedule its own follow-up weeks later.
The two decisive axes here are lasting memory and multiplayer. Because it permanently remembers a channel's instruction -- watch only for this type of issue, not those -- you never have to re-prompt; and a tool you once used one-on-one moves into the middle of the team, where several people can nudge a single session toward a better result.
How Does Adoption Spread on Its Own?
The most striking part is how it spreads. Because AI runs in public channels, everyone can see how the best users phrase their instructions and hand off work. People observe those patterns, carry them into their own projects, and best practices diffuse across the organization by themselves -- something rarely seen with earlier AI tools.
The effect shows up in numbers too. In the conversation, roughly 65% of pull requests (code change requests) in that product org are already written this way, and the share keeps climbing.
From an AX View, What Must an Organization Put in Place?
The conclusion I stress in AX work repeats here: results are decided by environment design, not by a smarter model. A public channel for AI to live in, a connection to source-of-truth documents that ground its answers, and memory that holds each channel's instructions. With those three, even non-engineers can contribute to the codebase without opening a terminal, and the questions new hires used to ask Legal or HR get answered instantly by AI that references the source of truth.
So the gap of the next year comes down not to which model you pick but to one decision: keep AI as a personal productivity tool, or make it shared infrastructure the whole team climbs onto. Organizations that choose the latter don't just change tools -- they change how they work.