One person working with one agent is pretty much figured out. Pick your tool, describe the work, get the work. The unsolved part is the team: several people, each with their own agent and their own sessions, trying to build one thing together. The answer that has emerged is shared context. Put the team’s history somewhere every session can read it, and the sessions stop being strangers to each other.
The amazing part
Shared history is the biggest jump I’ve seen in how teams can work. What I learned in my session yesterday, you inherit this morning. What your agent figured out last week, mine builds on today. A new teammate doesn’t need a month of catch-up conversations, because the decisions are written down where the tools can find them. Every tool got to the same shape from its own direction: a file of directions and settled decisions that each session reads before it starts, plus the memory the sessions leave behind.
I run even my document work this way now. The team presentations I’ve struggled with for years live in one shared place that everyone’s tools can read, and it has taken multiple rounds of building and rebuilding to get them right. Most of the people I make those documents with don’t know git and shouldn’t have to. The point is the shared place, not the plumbing.
This is what I mean when I say the real promise of agents is collaborative development. Not one person going faster. A team finally holding one history.
The risk
The risk lives in the same machinery. A team’s shared history is only as good as what the team agreed to. Anthropic’s 2023 sycophancy study found that models trained with human feedback learn that agreement can score better than accuracy — in people’s ratings, convincingly-written agreement sometimes beat correct answers. Every session carries that pull. Now let the sessions feed one shared file. A decision gets written down in March. By October it reads like fact. Nobody reopens it, the file says it’s settled, and every new session builds on it helpfully. That is how a team ends up wrong together with everything looking agreed.
There’s a non-zero chance that some of what accumulates in the shared history is agreement rather than fact. Not every direction in the file deserves to be agreed with. Someone on the team has to be allowed to reopen a settled question, and the setup should make that easy instead of arguing against it.
The book this site accompanies is the one-person version of this argument: what happens when the voice you talk to is tuned to agree (chapter 5), and how to take charge of what these systems remember (chapter 8). The team version is the same problem with more people in it. Most teams are already running it.