Skip to main content

Understanding Results

Start with analysis-report.md in your output directory. It explains proposed next steps, why they matter, and how to try and evaluate them. Jejoo proposes changes; running an analysis does not modify your agent.

Findings and recommendations​

A finding is an observation worth investigating, such as repeated tool execution, concentrated latency, or output-budget failures. A recommendation is a proposed response supported by the investigation. A finding may lead to a change, an experiment, a request for more information, or no action.

The customer report focuses on final recommendations. It does not list every intermediate detector result. Findings with a recorded no-action decision may appear under Findings with no action proposed, with the reason no intervention was selected. The full result and audit information are in analysis-result.json.

Report sections​

Section or labelHow to read it
Suggested next stepsChanges or checks for the observed workflow.
Architecture proposalsProposed structural changes that require validation before adoption.
Recommended changeAn intervention supported by available evidence; still test it on your workload.
ExperimentA hypothesis to evaluate through a controlled trial.
More information neededA request for evidence before choosing an intervention.
DeniedA change that was rejected; do not treat it as an instruction to implement.

Each item may include Why, Workflow targets, How to try it, What remains uncertain, How to check it worked, and estimated impact, quality risk, or implementation effort. Available fields depend on the evidence and type of recommendation.

Example interpretation​

Suppose a report identifies two searches with the same query and proposes reusing the first result. Check the cited operations and the explanation before introducing a cache: the second call might be intentional, require fresher data, or run with different permissions. A safe trial could compare tool calls and end-to-end latency while checking answer correctness and freshness.

This is illustrative, not a promised output from the sample dataset.

Estimates are not measured improvements​

Cost and latency savings are estimates tied to the analyzed workload and its available measurements. Missing estimates mean the result does not provide a supported numeric estimate. Do not interpret an absent value as zero savings, multiply an estimate to annual savings without a workload assumption, or add overlapping estimates together.

Risk, effort, and confidence help prioritize review. They do not replace testing. A faster model response is only useful if it still satisfies the task's quality requirements.

Output files​

FilePurpose
analysis-report.mdHuman-readable recommendations and validation steps.
recommendations.jsonCustomer-facing structured projection with final_recommendations.recommendations and finding decisions.
analysis-result.jsonJob status, evidence audit, budgets, and final analysis under result.
normalization-usage.jsonNormalization attempt counts, token usage, and estimated cost when available.
declared-workflow.jsonSupplied normalized workflow or locally inferred observed workflow.
receipt.jsonSaved job/submission metadata used for recovery.
local/manifest.jsonLocal fact revision, identities, and hashes.
local/facts/execution_facts.parquetCanonical execution facts retained on your machine.

Normalization cost measures the work of preparing the analysis, not the cost of the agent being analyzed. Cached usage is previous usage, not a new model call. Reasoning tokens are included in output totals; do not count them twice. Usage cost estimates are not invoices.

Protect the output directory as trace data. It may contain source exports, local paths, and sensitive content. The CLI currently saves local results rather than printing a shareable web report link.

Validate a recommendation​

  1. Read its evidence, target operations, assumptions, and remaining uncertainties.
  2. Record the baseline: cost per successful result, end-to-end latency, failure rate, and task quality.
  3. Try one coherent intervention on a representative workload.
  4. Compare the same metrics, including retries and failed attempts, against the baseline.
  5. Keep the change only if it meets your quality and reliability criteria. Analyze the new traces in a fresh output directory.

If the report has no recommendations, check the job status and evidence coverage. A completed run with no supported intervention differs from a failed or incomplete analysis. Missing workflow context or sparse traces can limit conclusions; a no-action explanation can also be a valid result.