All posts
Engineering
·
12 min read
Tracing multi-agent systems without losing your mind
Handoffs, shared memory and parallel workers turn a simple trace into a graph. Here’s how we model it so a human can still read it at 2am.
Jonah Weller
Founding Engineer
From trees to graphs
A single agent produces a tidy tree of spans: one root, a handful of children, a clear beginning and end. Add a planner that delegates to three workers, each with their own tools and retries, and the tree quietly becomes a graph with shared state flowing between branches.
We tried rendering that graph literally. Nobody could read it. What engineers actually want during an incident is the critical path — the sequence of steps that determined the outcome — with everything else folded away until they ask for it.
Critical path first
Sable computes the critical path for every run and renders it as the default view. Parallel branches collapse into a single row with a count, and handoffs between agents become explicit edges carrying the exact context that was passed along.
The result: our median time-to-root-cause for multi-agent incidents dropped from 41 minutes to under 6.