01
Blog
Field notes on shipping agents.
Evals, tracing, cost and the things that broke. Written by the people who build Tracewell.
Product
Shipping v2.4: evals as a deploy gate
A version now ships only when its evals clear the bar you set. Here is how it works.
Guides
Evals are tests. Treat them like it.
Version them, review them, run them in CI and let a failing one block a merge.
Guides
The anatomy of a trace
Spans, attributes and the three numbers worth reading first on any run.
Engineering
Why we record tool calls, not just model calls
Half of the bugs we see in agents are in the tools around the model. Tracing only the model hides them.
Guides
Canary releases for agents
Send five percent of traffic to a new version and let the evals watch it.
Engineering
A refund bug, a four-minute fix
How a support team at Halcyon found the prompt that broke refunds, and what it replaced.
Company
What we learned from 12 million runs
Patterns across a year of traces: where agents spend time, money and trust.
Guides
Writing a rubric a model cannot game
Specific beats clever. Four habits for rubrics that keep working after the model gets better.
Engineering
p95 is a lie: tail latency in agent loops
Loops multiply latency. A look at why the slowest one percent of runs matter more than the average.
Company
How we price: runs, not seats
The person who needs to open a trace is rarely the person who bought the tool.
Engineering
Redacting PII before it leaves your server
Redaction rules run in the SDK, so sensitive fields never reach our servers.