Skip to main content
Improvement starts after a run has been inspected and the failure mode is clear. Sovara helps you turn that review into a change that is specific enough to fix the problem without making nearby cases worse. The important decision is where the fix belongs. Some failures need code. Some need a better prompt. Some need a new tool. Some come from domain knowledge the agent lacks, where the right fix is a reusable prior in SovaraDB.

Choose the right fix

Use the smallest fix that addresses the actual failure:
Failure patternBetter fix
The agent could not perform an actionAdd or improve a tool
The agent did not follow instructions that are not context dependentUpdate the prompt or workflow
The agent missed a domain rule or context-dependent instructionCreate a SovaraDB prior
The goal is not to turn every annotation into a prior. A prior is useful when the lesson is reusable, conditional, and likely to matter again. Learn more in What are priors?.

Make domain lessons reusable

When the failure comes from missing domain knowledge, open SovaraDB and create a prior. Capture the lesson in a form that is precise, scoped, and retrieval-friendly. Click New Prior, write the rule, and use Suggest to draft the Use this when field. Before relying on the prior, review it with Review. The detailed workflow is covered in Creating and managing priors.

Validate the change

After applying a fix, rerun representative examples and inspect the new traces. The question is not only whether the original failure disappeared. Check whether the fix changed related behavior in a way you did not intend. Use Runs to compare executions, Manual inspection to inspect the changed steps, and SovaraChat to get a fast read on the new run. For failures that should become regression tests, use Benchmarks to create repeatable coverage.

Close the loop

Once a fix works, keep the annotation history. It tells Sovara which behavior is already covered and helps the recommendation algorithm avoid wasting reviewer time on examples that no longer teach anything new. If similar failures keep appearing, the fix is probably too narrow, not being retrieved at the right time, or applied at the wrong layer. In that case, go back to the reviewed runs and refine the prompt, tool, prior, or benchmark until the pattern is covered.