Welcome to Sojourner Labs

Plenty of people are building multi-agent systems. Very few are measuring them carefully. Most claims about agent teams come down to “it felt better,” and I want numbers instead.

I build Sojourner, an orchestration engine where agents work like a software team, with a planner, specialists, a reviewer and a tester. This site is where I take that engine apart and measure which pieces actually earn their cost.

The questions

These are the questions I’m starting with:

• Team structure. Does a manager agent directing specialists beat a flat team of peers?
• Review loops. How much does a dedicated reviewer improve the result, and does a second one add anything?
• Model mixing. Which roles can run on cheaper models without hurting quality?
• Failure and recovery. When one agent goes off track, how often do the others catch it?

How I’ll measure

Every experiment runs a fixed set of benchmark tasks, with small apps and features of varying difficulty. I change one variable at a time and record:

Metric	What it captures
Pass rate	Does the output meet the task’s acceptance tests?
Cost	Total tokens and dollars per task
Wall time	How long the run took from start to finish
Interventions	How often the team needed a retry or a human nudge

Raw results and configs live in the same repo as this site, so anyone can check my work.

What’s next

The first experiment compares a manager-led team with a flat team on the same tasks. Results soon.