Your agent doesn’t need more context, it needs fewer choices
I cut one agent’s tool list from nineteen down to six, and it got better at all six. How I decided what to take away, and the one tool I was wrong to remove.
The model is the easy part. Everything that makes an agent safe to rely on sits around it, and almost none of it is interesting enough that anyone writes it down. So I’m writing it down.
moo.fm is being built mostly by agents. This is what trips people up, and the boring practices that take you from an agent you hope works to one you can trust.
It’s a journal more than a blog. No schedule. It turns up when it turns up.
I cut one agent’s tool list from nineteen down to six, and it got better at all six. How I decided what to take away, and the one tool I was wrong to remove.
A pipeline that works most of the time is a pipeline you can’t debug. What I pinned down, and what I gave up in order to pin it.
The blast radius is the design. Where I drew the line around credentials, customer data, and anything that can send mail.
I keep every prompt, tool call and result, and I can replay any run. It took a weekend and it has paid for itself twice over.
A hundred green evals, and not one of them checked the thing that mattered. What went into the suite afterwards.
· Evals
Adjectives in a prompt do almost nothing. Structure does almost everything. Three rewrites, with the before and after.
· Practice
Most bad output is a model with no exit. I built one in, and the confident nonsense mostly stopped.
· Agents
Twice I automated the wrong half. What finally made it hold, and how I could tell the difference.
· Automation
I made my agents narrate their decisions. It cost about four percent in speed and saved me weeks.
· Observability
It started as a complaint about my own calendar. The version I almost built instead, and why it was wrong.
· Practice