Evals & Lifecycle
Backtest every change
before you deploy.
Keep testing as you improve your agent team. Compare each candidate version against past work, inspect regressions before deployment, and restore a proven team version if a change underperforms.
Revision reviewWhat changed in the team?
One proposed assignment
v12 → v13Model A
Saved team revisionModel B
Test before adoptionChoose representative work
4 reviewed casesCompare outcomes by case
3 accepted in each runHold the current revision
review requiredIt passed the review exception in this example.
Inspect the case regression before another test.
The test set
Include the cases that can change the decision.
Corin’s retained history gives you past tasks to revisit, including routine work and exceptions. Keep each case’s inputs and reviewed outcome together where one exists.
Backtest as you go: run the current and candidate team versions against the same cases as you refine the configuration. Compare reviewed outcomes, time, tokens, and cost, and define what counts as a regression before deploying the change.
Four different paths through one task.
Routine, incomplete, exception and urgent work belong in the same comparison.
Same aggregate. Different failure.
Reviewer accepted the exception recommendation.
Reviewer asked for a correction on the same case.
Read the regression
Find the case that got worse.
Two revisions can have the same number of accepted cases while failing on different work. The reviewed exception, not the average, determines whether this candidate is ready.
Open the traces for the case. Compare the model choice, inputs, tool calls and handoff before changing the team again.
Version and restore
Restore a proven team version.
Corin retains telemetry and relevant data after each task ends. The record shows what changed across revisions and how the team performed, so you can inspect a deployed change and restore an earlier whole-team revision.
This example changes one model, but the saved revision covers the whole agentic team. A failed candidate stays out of use; a later adopted revision can be restored to the prior team configuration if it regresses.
Hold the current revision.
Do not replace v12 with v13 while the reviewed exception regresses.
Decision record
Keep the comparison explainable.
What changed
Identify the team revisions and the proposed configuration change.
What was tested
Keep the representative inputs and review criterion comparable.
What people decided
Save the reviewed outcomes, trace links and reason to adopt, hold or restore.
Build with Corin
Try a change on known work.
Run a working example, inspect its trace, and use representative cases to assess the next team revision.