JD FETTERLY / SELECTED WORK

Selected work

Some of the work behind what I write here, from changing how a delivery team works with agents to building them into my own day.

Delivery team

Rebuilding how a delivery team works

Illustration from the related essay.

When I asked my boss to show me how we were managing projects, he opened a customer’s project plan. Then another page for the next customer. Then another. He knew what was happening across those projects, but we needed a way for the team to see the commitments we were managing and where information was missing.

I built a pipeline that brought the projects into one view, including the dates that hadn’t been filled in. That gave us a starting point. My boss had already built the foundations of our delivery approach, and we worked together to refine how the team managed the work as we brought agents into it.

After customer meetings, delivery leads were piecing together notes and updating our customer and project records by hand. We connected the meeting types to the work expected at each stage of a project, so agents could propose updates and identify tasks to add or close. The delivery lead still had to decide whether those actions served the customer’s goals.

Preparation was another part of the same effort. We built a shared skill that assembled current materials and customer context into a presentation. I designed a lightweight registry for the team’s shared skills, with team leads owning their definitions. The person taking a presentation into a meeting remained responsible for its accuracy and for making it useful to the people in the room. Having a deck ready didn’t mean that work was finished.

Adoption took work too. Some people continued making updates manually, and information the agents knew to capture was still getting missed. Seeing the workflow run for a colleague helped people understand what they could hand over.

As preparation and follow-up took less time, we had more room for customers and for developing the team. Delivery leads could present their work to colleagues and get feedback before taking it into a customer meeting. Senior leads had more time to help them build the judgment they still needed when the agents had finished their part.

What are you automating? follows the meeting workflow. When is the work actually done? looks at preparation, shared skills and the responsibility that stays with the person.

Personal workflow

A team of agents around my commute

Illustration from the related essay.

I built a team of Grokbots around my commute because I didn’t want to keep managing all the small tasks involved in getting to the office. One checks my calendar and asks whether I’m driving. Courtney finds parking, and another agent handles the next step in preparing it for reimbursement.

The handoffs also depend on me. Courtney held a parking reservation while she waited for my approval. I took too long and the hold expired. She tried again, and I took too long again. She made three attempts to get parking arranged for the next morning.

Working out what the agents should do included deciding when they needed to come back to me. In that case, they were waiting on the person who had asked them to take the work off his plate.

After a month of having them manage the commute, I wasn’t sure how I’d got by without them. The bots don’t always get it right, but neither do I.

Read Grokbots manage my commute.

Builder project

Codex Artifact Gateway

I wanted a way to review locally generated HTML from my phone. Codex Artifact Gateway is a small project for that part of working with agents: opening their output and leaving feedback on it.

The implementation serves HTML and supporting files from a Mac over a private network, provides a recent-artifact view and saves feedback locally. It depends on the Mac and service being available. The public repository contains the implementation and tests; it’s an independent project, not an OpenAI product.

Explore the repository