Skip to content
XLinkedIn
Sign Up →

Findings

A Finding is not a single bad trace. It’s a recurring failure pattern across multiple sessions, with a root cause explanation and a proposed fix.

In the console, the Findings page groups by agent on the left. Select an agent to see its Findings on the right.

Findings page

FieldDescription
TitleShort name for the failure pattern, e.g., “Temporal / chronology confusion”
DescriptionOne-sentence root cause, e.g., “Model focuses on incident type rather than liability context”
TagsFailure category labels: missing_edge_cases, missing_context, vague_instructions, OPAQUE_TASK
Session countHow many sessions show this pattern. Higher = more impactful
First appearedWhen Kelet first detected this pattern
Prompt PatchProposed change to your agent’s system prompt

A Prompt Patch is a fast, single-shot proposed edit to your agent’s system prompt that addresses the root cause. Generated immediately by an LLM from the Finding’s context.

Apply it as-is, customize it, or ignore it.

For a validated fix — one tested against your real session data using an evolutionary search — run GEPA Optimization from the dropdown. Slower, but comes with a measurable quality delta. See GEPA Optimization.

Each Finding card has two action buttons:

  • Prompt Patch (split button) — view and apply the suggested fix. The dropdown also exposes GEPA Optimization ( re-run the evolutionary process), Mark as Resolved, and future options (disabled for now).
  • Not an issue — mark as false positive; removes the card from the Open view.
  • Open — active Findings that haven’t been resolved or dismissed
  • Resolved — Findings where the fix has been deployed
  1. Go to Findings in the console
  2. Select your agent on the left
  3. Sort by session count — higher count = more impact
  4. Read the description — is this a real problem you recognize?
  5. Apply or dismiss

Findings improve over time as more sessions accumulate. A pattern seen in 3 sessions is a hint; a pattern seen in 50 sessions is a root cause.