What is Tensile
Tensile turns agent traces into evidence-backed evaluators and a review workflow for improving reliability.
Watch the workflow: Inspect authored workflow evidence from Home · Authored dummy data
The video shows the earlier Home navigation. Migrated workspaces open the same read-only study from Investigations.
See one evidence-reading step on the Tensile Home page using authored demo data. This clip does not demonstrate every product capability or approve an evaluator. Filter authored examples, inspect eligible evidence counts and follow a saved chronology without claiming a customer outcome.
Read the walkthrough transcript
- 00:00 — Filter the Tensile study to authored examples. Customer outcomes remain unverified.
- 00:05 — Open a pattern. Compare eligible counts and the evidence shown.
- 00:11 — Open Chronology. This invented attempt has no saved completion.
Tensile helps teams understand and evaluate how AI agents behave in real workflows. Connect trace sources or instrument an agent, then use its evidence to define what success and failure mean for a repeated business decision.
The product brings workflow discovery, evaluator-pack authoring, evidence review, and reliability analysis into one workspace. Configured organizations can also compare a declared task with destination evidence, assess whether human review can change, and inspect a scoped baseline-versus-candidate comparison. Trace data is evidence for review; it does not make a draft evaluator approved or cause it to score live traffic automatically.
Find a workspace
Open Observability for Pulse, Error Captures, Agent executions and Analytics. The sections keep their existing filters, saved views and record links. Environments opens controlled environment tests; provider setup remains in Settings. User Stories remains a separate destination.
Open the book icon at the right of the app header for Docs. It is also available in Settings and opens documentation in a new tab.
The reliability workflow
Bring in traces from supported integrations or send turns and tool calls through the ingestion API.
Discover repeated workflows and draft versioned evaluators that make the business decision and evidence explicit.
Review real examples and source evidence. Calibration, independent validation, and owner approval gate any production use.
The Reliability scorecard opens with the last seven days. Use its date filter to choose a preset or custom interval, then select an environment and customer. Open a metric’s short detail line to see its counts and coverage limits. A missing measurement remains unknown; a trace signal or evaluator verdict does not establish a verified customer outcome.
How the signal flows
Tensile can receive new events from the ingestion API or import existing traces from a supported connector. The exact data available depends on the source and its configuration.
- Instrument
Connect a supported source such as Braintrust or Langfuse, or send your own turns and tool calls through the Tensile ingestion API.
- Ingest
Tensile normalizes source records into trace sessions and events. Events retain their parent relationships and source identity when the source provides them.
- Distill
Workflow discovery and evaluator authoring use the available trace and source evidence. Reviewers see the question, answer criteria, and supporting evidence for eligible cases.
- Improve
A draft pack is not a live score. Calibration, sealed validation, and authorized owner approval must pass before a pack can be promoted for production use.
Start here
Connect a source or send an event, then follow the steps to check that Tensile can read its trace data.
How to keep user data pseudonymous, allowlist metadata, and redact sensitive content before it leaves your app.
Learn how packs move from discovered workflows and draft definitions through human review and validation gates.
Follow an Outcome case from the agent's claim to destination evidence, then inspect Readiness and Changes.
Connect a Braintrust or Langfuse workspace and understand what a trace connection does.