Telemetry & Optimization
See what happened.
Find what needs fixing.
Debug with detailed agent and system logs from built-in Langfuse. Follow each run through its context, model choices, tool calls, and handoffs, then backtest proposed changes on representative past cases.
Telemetry explorerFollow the work, step by step.
Trace waterfall
Nested work reveals where time accumulates.
Token use by role
Input and output are separate for each agent.
Cost by model assignment
Trace costs back to the roles that incurred them.
share
Reviewed cases
Keep outcomes attached to representative work.
History across tasks
Detailed logs make debugging easier.
Corin retains agent and system telemetry across tasks: which agent and model took a step, which tools were called, where work was handed off, and the tokens, latency, cost and measured outcomes.
Inspect a run step by step to locate an unexpected result, a failed tool call, or a slow handoff. The retained trace connects the result to the agent, model, and tool steps behind it, so you can investigate what needs fixing.
Same cases, different model.
Compare the output and reviewed result before accepting a lower-cost assignment.
Evaluate a change
Backtest a candidate before switching.
For one agent, change the model assignment and run the candidate on representative cases. Compare token use, latency and cost with the current setup.
Where reviewed outcomes exist, compare answer quality too. A lower cost or faster response does not establish that the recommendation is better.
Agent and system view
Measure work across agents and handoffs.
One agent's trace explains its own steps. System telemetry follows the work across roles, tools and handoffs, so a team change can be assessed against the full task.
In a quote task, speeding up the pricing step may barely change the full task if review takes most of the time. Compare task time, usage and reviewed outcomes where available before changing roles or models.
Which role takes the time?
This step contributes 7.4s in the sample task. Open its trace to see the handoff and decision.
Read the evidence carefully
Usage is measured. Quality needs a reference.
Tokens, time and cost are observable signals. Accuracy requires a defined task, representative cases and outcomes that can be reviewed. If a case has no reliable outcome, mark that limit rather than scoring it as a success.
Inspect the run
Find the agent, model, tool call or handoff behind the result.
Compare the same work
Use representative cases for the current setup and a candidate change.
Review what mattered
Use a human-reviewed result where one is available.
Build with Corin
Run a task. Inspect its trace.
Start with a working example, examine the agent and team steps, then choose a change you can test on relevant cases.