When agents join us

I remember when I first began to think that AI would change how we work with each other. It was February of this year, and I was watching a friend use Claude Code to update articles for SEO at his growth agency. It was the first time I’d seen an agent work directly on a website, recognizing elements on the page, interacting with them, and making changes through the terminal.

I immediately started imagining everyone working with their own agents, subagents, and skills. Then I began wondering how those individual ways of working would fit together.

When AI helps one person do more, what makes the rest of the company more capable?

That question has stayed with me through conversations with teams at wealth management firms, family businesses, startups, and IT consultancies. The OpenAI employees I spoke with described doing their coding and collaboration entirely through agents. In a separate conversation, a team described people completing work independently without much clarity about what their colleagues were doing.

I wanted to understand what others could learn from that work, and how it connected to the decisions the company needed to make.

I had worked on a version of this challenge at Yummy.

There, I helped a very small team of data champions get tools to build tables, data models, and metrics using the company’s own data. I also brought them into the rooms and teams where decisions were being evaluated and made.

Those two things belonged together. The champions needed the ability to build and the context to understand which problems mattered. Leaders could guide the direction, while people who understood the data helped the company evaluate its options.

Their capability became more useful because it had a place in the company’s decisions.

That experience shapes how I think about AI. Giving someone a powerful tool creates possibilities. For colleagues to benefit, they also need to understand what that person built, what it relies on, and where it could help.

A funny video from the founder of Leadsales gave me another way to picture this. He was encouraging people to use public Slack channels instead of private messages. It made me imagine people and agents sharing unfinished work, asking questions, and reviewing things as they developed.

This is a setup I still need to explore. I can imagine overlapping conversations and competing suggestions. I can also imagine a colleague catching a missing detail early or recognizing that another team has already solved part of the problem.

That gives visibility a purpose: helping people contribute to the work and carry something useful away from it.

This is also why I agree with Willie Williams’s distinction, described by Laura Entis in Every, between a company agent and one “that is just yours.” I’d want my personal agent to grow with me as I learn and figure things out, independently of where I work.

At work, I’d want my colleagues to be able to pick up a useful workflow or understand a decision without needing access to my personal bot. I’d still want room to think privately, including being uncertain or changing my mind. Making the work easier to continue shouldn’t require sharing every thought that went into it.

What they can see matters, too. A metric or an agent’s output can look convincing while relying on incomplete data or assumptions nobody has questioned. People need access appropriate to their responsibilities and enough context to judge whether the work applies to their situation. Someone still needs to own the outcome and make the final call.

Over time, I would hope to see useful work become easier for others to build on. A team could reuse a skill, ask a question earlier, or revisit the reasoning behind a decision. The next project would begin with something the previous one taught them.

Sometimes that learning would improve a process. Sometimes it would reveal that the process needs to be rebuilt.

Each company will have to develop its own practices for this. Outside specialists can help, and their work should leave people inside the organization more able to understand, evaluate, and improve their systems themselves.

Success will look different in each company. Effective AI implementation should solve a painful, time-consuming problem, with improvements visible in the work: less time lost, fewer repeated mistakes, or clearer decisions. A team that is already effective may benefit more from protecting its focus than introducing another system.

That is what I am exploring at Reframed: how work done by one person can leave a whole team more capable.

When the next person picks up the work, they should have a better place to start.