When a traditional API fails, you get a stack trace. When an agent gives a bad answer, the cause could be a vague prompt, a missing document, a failed tool call or a model that simply guessed. Observability makes the difference between guessing and knowing.

Trace the whole run

A good trace shows every step an agent took: the prompt, the model's reasoning summary, each tool call with its inputs and outputs, and how long each step took. Codexa records this automatically for every sandbox run.

Measure what matters

• Task success rate against your evaluation set

• Median and p95 time to complete a task

• Tokens and compute cost per task

• Tool error rate by tool

Debug faster

Filter traces by outcome, replay a failed run in a fresh sandbox from its snapshot, and compare it side by side with a successful one. Most issues become obvious within minutes.

Start with tracing on every run and a weekly review of your ten worst traces. It's the fastest way to make an agent better.