Blog · AI at work
The 20 AI agents that run my workday
I manage a program inside one of AMD's biggest AI hardware platforms. A lot of my week used to go into work that kept the program moving but didn't need my judgment. Now about 20 AI agents do most of it.
Why I started
Program management at this scale is mostly coordination. Many teams, many dependencies, many dates. Before I could make a single decision, I had to gather status from different sources, reconcile plans that had drifted apart, and turn it all into updates for different audiences.
None of that was hard. It was just constant, and it crowded out the part of the job that matters: spotting risks early and helping teams make good calls. So I started handing the repetitive parts to AI, one task at a time.
What the agents do
I can't share the specifics of the program, but the agents fall into a few broad groups.
🗓️ Planning
Keep plans and schedules in sync across teams, and flag when they start to drift.
📈 Forecasting
Look ahead at schedule and resource needs, so surprises show up weeks earlier.
📣 Reporting
Pull status from many sources and draft the updates each audience needs, from engineers to leadership.
✅ Follow-through
Turn meetings and threads into action items, and keep track of who owes what by when.
They're built with Claude, using custom skills, plugins and Model Context Protocol (MCP) integrations that connect them to the tools the program already runs on. Each one is small and does one job well. Together, they behave like a support team that never sleeps.
What changed
Planning and reporting that used to take days now take minutes, and the program is moving ahead of its original schedule.
Work that once needed extra contract support is now handled by the agents.
With most routine work automated, I've been able to take on additional responsibilities.
What I've learned
- Start with the most boring task. The work you dread every week is usually the easiest to automate and the most rewarding to hand off.
- One agent, one job. Small, focused agents are easier to trust, fix and improve than one agent that tries to do everything.
- Keep a human on the decisions. The agents gather, draft and flag. I review, decide and own the outcome.
- Write instructions like you'd brief a new teammate. Plain-language skills are easy for anyone to read and improve.
The goal was never to automate my job. It was to automate the parts of it that kept me from doing my job well.