Building better software by day, chasing sports stats by night. Occasional writer about both. Always learning and evolving.
Minnesota, USA
Joined March 2023
- Tweets238
- Following195
- Followers31
- Likes10
Pinned Tweet
Getting an AI agent to do a task is easy.
Keeping its work coherent for weeks - when it forgets decisions, idles invisibly, and grades its own work - is not.
I run agents on long-running work. A pattern that works for me is clearer responsibilities instead of better prompts. 🧵
What reaches me from the research is flags and exceptions, each one a decision to make with a recommendation attached. I review those instead of every change. That is the escalation contract from my orchestration essay, in production: roughwork.dev/blog/agent-orc…
I built a public archive of NFL trades going back to 1920. A draft pick can be traded years before it is used and pass through several teams, so the archive traces every stop a pick made. It now holds roughly 4,000 trades and more than 5,000 draft picks.
A script checks every chain: each receiver is the next sender, the trades run in date order, and the chain ends at a real draft slot. What fails goes to a review queue. Flags that reach me carry a recommendation, and I rule on those instead of reviewing every change.
The coordinator agent and the research agents run on the same underlying model, so separate roles never added up to an independent second opinion. I wrote about the system, its checks, and what I can't verify about a trade from 1954 here: roughwork.dev/blog/century-o…
6. Treat human interventions as defects.
Every time you have to restart a worker, supply a missing decision, or catch an error - log it. Was a check missing (structural) or was a rule ignored (judgment)?
They look alike. They need different fixes.
7. Prune the apparatus.
Every failure tempts you to add a check. Checks accumulate and outlive their problems.
Keep, merge, or delete it. Each should name the failure it prevents.
Full essay, with limits and where to start:
roughwork.dev/blog/agent-orc…
1/10 After analyzing over 1,300 NBA seasons since 1976, I discovered something fascinating about when teams show their "true" identity.
9/10 But something has changed in the last decade. Since 2014-15, good teams show themselves much faster (16.07 games) while bad teams still need ~22 games.
10/10 The 2017-18 season was particularly weird. Teams needed an average of 27.5 games to find their identity. Why? Read on... buff.ly/3ULF16u