Anthropic
Anthropic builds AI systems that are safe, beneficial, and understandable. Their product teams ship agentic features that orchestrate long chains of tool calls — exactly the kind of system that's hardest to debug from a dashboard.
The challenge
Agentic features fail in ways that don't show up as a single bad request. A run might span dozens of tool calls, retries, and model hops before something quietly goes wrong three steps upstream of the error anyone notices.
The existing setup leaned on the usual three pillars — metrics, traces, and a log firehose nobody read until an incident. When an agent loop stalled in production, the on-call engineer was left scrolling spans, guessing which hop introduced the regression. Mean time to first hypothesis was measured in hours, not minutes.
The solution
Anthropic piped their agent runtime logs through Sazabi and started asking questions in plain language instead of building dashboards ahead of time.
Before
- Pre-built dashboards for known failure modes
- Manual span-by-span trace reading
- Tribal knowledge locked in a few senior engineers
After
- Ask "why did this run stall?" and get the offending hop
- Failure modes surfaced before they page anyone
- Any engineer can interrogate the stack on day one
We can finally interrogate our stack the same way we'd interrogate a model. The questions are the same — what did you do, and why? — and now the infrastructure actually answers.
Results
- 74% faster time-to-first-hypothesis on agent incidents
- 3× more production issues caught before a customer report
- On-call ramp for new engineers cut from two weeks to two days
