Carlos Granier
Back to homelab
Records & decisions ¡ Happened

The bottleneck was me

I assumed my agents were the bottleneck, so I counted: 37 of 51 open tasks were waiting on my decision.

Still true Accurate as of
Waffle chart of 51 squares, one per open task on September 27. 37 squares are highlighted as waiting on me; 14 are waiting on agents or others. 11 of the 17 high-priority tasks were mine, and the oldest had waited 48 days.

The short version: I run my projects with AI agents and assumed they were the slow part. On September 27 I counted: 37 of 51 open tasks were waiting on me, including 11 of the 17 high-priority ones. One had waited 48 days. So I built a way to make decisions faster instead of adding agents.

Get the week's seven in one email, every Monday.

What went wrong

Every unit of work in my setup is a small task record with an owner, a status and a next step. Agents pick up their own records; mine wait for me.

I kept thinking about how to get more out of the agents. Then I counted where the work was actually stuck. Most of it was stuck on me: approvals, choices, “should we do this at all”. An overnight agent that proposes work had suggested the same one-line fix on 15 of 16 nights. The agent wasn’t slow. My decision was.

What we learned

  • Measure where your queue waits before you add capacity. With agents, the queue that needs the most attention is often the one with your name on it.
  • A report-only agent moves the bottleneck to you. If the same proposal comes back night after night, you have a decision problem, not a compute problem.
  • Decisions need their own interface. Mine were scattered across chats, notes and records, so none of them felt urgent.

What went right

I built a decision desk: one Telegram card at a time on my phone, with buttons for approve, proceed, decide, defer, kill and skip. Every answer goes into a ledger of what was proposed and what I decided. The first walk-through took about six hours and cleared 41 decisions. 41% of them were skips, which is useful information too.

How to do it (as of October 2026)

  1. Count open work by owner, and by owner among the high-priority items. Do it before you build anything.
  2. Put decisions in one place you’ll actually open, one at a time.
  3. Log each proposal next to the decision. Over time, that ledger shows which kinds of work you always approve, and so which you could let run on their own.
  4. Make “skip” mean “later”, not “show me again right now”. My first version brought skipped cards straight back.
  5. Sort by priority before “someone is waiting”. Mine briefly put a minutes-old normal task above a 25-day-old urgent one.

Status

Still true. On October 12, 52 of my 97 open tasks were waiting on me. The desk helps when I use it; using it more often is the open part.

The week's seven, by email

Entries post here daily. Every Monday, one email with the past week's seven. Unsubscribe anytime.

Let's Connect