How Tensile works
From agent traces to evidence-backed evaluator packs, human review, and gated production use.
Tensile connects trace evidence to the decisions a team wants its agent to make reliably. The workspace helps people define those decisions, inspect real examples, and decide whether an evaluator has enough evidence to be trusted.
The path from traces to review
- Connect a trace source
Use a supported integration such as Braintrust or Langfuse, or send events with the SDKs or ingestion API. What Tensile can use depends on the source and its available fields.
- Reconstruct the workflow
Tensile groups related trace events and uses available event, tool, and state evidence to identify repeated workflows, decisions, and possible outcomes. Missing or conflicting evidence stays visible as uncertainty.
- Draft a versioned evaluator pack
A pack defines the workflow it applies to, the decision or outcome being evaluated, the evidence recipe, and the answer criteria. Proposed fields and rules retain their sources and support status.
- Review examples
An authorized reviewer sees the plain-language question, answer meanings, decisive evidence, and source provenance where available. The reviewer can save a draft, lock an answer, and move through eligible examples. Model answers remain hidden until the human answer is locked.
- Validate before production use
Calibration, an untouched sealed validation set, independent review, and authorized owner approval are separate gates. A draft or shadow result is not an official score and does not automatically trigger routing or other actions.
What is stored where
- Trace data contains imported or directly ingested sessions and events. The source, execution or trace identity, timestamps, parent relationships, and other provenance are retained when the source provides them.
- Evaluator state contains organization-scoped workflows, pack versions, evidence recipes, review answers, judgments, and approval history.
- Model output may help propose findings, evaluator definitions, or judgments. It is recorded against the relevant inputs and version; by itself, it is not a human label, customer policy, or approval.
Why versions and evidence matter
An evaluator's question, answer criteria, or evidence recipe affects the judgment it produces. Tensile keeps packs versioned so a material change can be reviewed as a new definition. Evidence shown to a reviewer is tied to the trace/source record where that information is available. A customer action or model score is not silently promoted into a correctness label.
Current product boundary
Connecting traces makes them available to the configured workspace. It does not automatically approve a pack or enable continuous scoring for every organization. Production use depends on the pack's validation state and the organization's authorization. See Evaluator packs for the review lifecycle and the data handling guide before sending trace content.