Tracing multi-agent systems without drowning in spans
Marcus Osei ·
A twelve-step agent run produces hundreds of spans, and a trace view that shows all of them is as unreadable as the logs it replaced.
What we changed
Collapse retries into the span they retried, with a count.
Record tool inputs and outputs, but truncate them — the first 500 characters answer most questions.
Give every run a step budget so an infinite loop is a alert rather than an invoice.
The full walkthrough is in the documentation.