01
Wrong tool call
Agent selects refund_order() instead of check_order().
Autonomous QA for AI agentsEarly accessWe try to break your AI
We try to break your AI
before your customers do.
Your agent passed the eval. Can it handle the customer?
We simulate real users, catch what changed, and turn agent failures into bugs your engineers can reproduce.
Staging-firstNo production dataNo SDK to get started
THE RELEASE CHECK YOUR AGENT IS MISSINGILLUSTRATIVE PRODUCT PREVIEW
VERSION COMPARISON
Agent v1.8 Agent v1.9
324 scenarios completed · Identical scenarios. Two versions.
Task success+2 pp
94%96%
Baseline → Candidate
Tool accuracy−15 pp
97%82%
Baseline → Candidate
Policy adherenceNo change
98%98%
Baseline → Candidate
Avg latency−1.9s
8.1s6.2s
Baseline → Candidate
11 new regressionsA better overall score. A broken booking flow.Inspect failures
NEW IN CANDIDATE 11
Duplicate bookingSCN-037 · State management
Unrequested refundSCN-084 · Tool selection
False confirmationSCN-112 · Tool failure
Critical regressionSCN-037
It said “rescheduled.” It booked twice.
EXPECTED
Tuesday Cancelled
Wednesday Active
ACTUAL
Tuesday Active
Wednesday Active
Run completeBaseline preserved/Reproducible failures attached
Sample data to show the workflow. Your audit is built around your agent.
BUILT FOR AGENTS
THAT TAKE ACTION
THAT TAKE ACTION
Support
Sales
Workflows
Browser & coding
RAG & multi-agent
01 / THE BLIND SPOT
A small change.
A very different agent.
A new model. A prompt edit. A tool update. Your test suite stays green while your agent starts doing the wrong thing.
We test behavior across full conversations, where retrieval, memory, and orchestration can quietly break.
02
Lost state
The user's goal disappears after several turns.
03
Duplicate action
A new booking is created without removing the old one.
04
Hallucinated success
The agent confirms completion even though the tool failed.
05
Long-conversation failure
Behavior degrades as the conversation grows.
06
Cost / latency regression
A model swap makes every task 3× slower.
Correct answers are only part of the job. Correct actions are the rest.
02 / THE AUTONOMOUS QA LOOP
Your staging API.
Our most difficult users.
Tell us what your agent should do. We explore what happens when users change their minds, tools fail, and conversations get complicated.
POST /agent
{
"message": "Actually, change my booking.",
"session_id": "synthetic-user-037",
"environment": "staging"
}Agent connected
SYNTHETIC USERS
Changes their mindGives conflicting inputAsks the unexpectedStays for 30 turns
A bug your engineer can act onScenario + replay + expected vs. actual
- 01
Connect your staging agent
Provide an endpoint, schema, and a short description of what your agent should do.
- 02
Generate scenarios
We build realistic behavioral scenarios around its capabilities and constraints.
- 03
Simulate users
Synthetic users interact across multi-turn conversations—not just one-shot prompts.
- 04
Detect regressions
Identical scenarios run against baseline and candidate versions.
- 05
Replay the failure
See the exact conversation, tool calls, expected behavior, and failure reason.
03 / FROM SCORE TO ROOT CAUSE
Don't just get an eval score.
Get the bug.
A higher task-success score can hide a new regression. See the conversation, inspect the tool calls, and understand exactly what failed.
Find out what breaks your agentSCN-037 / State management
CriticalUser reschedules an existing meeting
CONVERSATION REPLAY 4 turns
TOOL EXECUTION TRACE
create_booking(tuesday)successcreate_booking(wednesday)successcancel_booking(tuesday)missingExpected
Tuesday bookingCancelled
Wednesday bookingActive
Actual
Tuesday bookingActive
Wednesday bookingActive
CONSTRAINT VIOLATED
Rescheduling must leave exactly one active booking.
The agent confirmed a reschedule but never cancelled the original booking. Two meetings remain active.
Baseline vs. candidateExpected vs. actualRepeated verification
04 / LESS TEST AUTHORING. MORE FINDING BUGS.
You build the agent.
We think like the user who breaks it.
Eval tools help you run the tests you write. Tensile is being built to discover the behavioral tests you haven't thought to write yet.
TEST EXECUTION
The traditional eval workflow
- Imagine every possible failure
- Write test cases and prepare datasets
- Build scorers and maintain the suite
- Investigate what a lower score means
EngineerTestsEval platformScore
BEHAVIOR EXPLORATION
Tensile
- Describe your agent and its constraints
- Generate scenarios and synthetic users
- Explore behavior across agent versions
- Receive failures with evidence and replays
Your agentAutonomous QAReproducible bug
We're building an automated QA engineer for your agent.
Less time inventing test cases. More clarity on what to fix.
START WITH A FREE AUDIT
Let's find your agent's
first “how did we miss that?”
$0
Your first agent QA audit.
A concrete place to start.
We're inviting early AI teams to test real staging agents with us. Share what your agent should do. We'll run a focused synthetic audit and walk you through the failures we can reproduce.
You provide
- Staging agent endpoint
- Dedicated test credentials
- Request / response schema
- Expected behavior and constraints
We investigate
- Realistic user personas
- Multi-turn conversations
- Edge cases and tool failures
- Repeated failure verification
You receive
- Reproducible failure scenarios
- Expected vs. actual behavior
- Severity and reproduction rate
- Conversation replays
No production credentials required.
Start with staging.
Keep production out of it.
Staging-first
Initial testing happens against staging or test environments.
Test credentials
Use dedicated test accounts, never production credentials.
No customer data required
Your first audit needs no real customer conversations.
Black-box friendly
Start through your existing agent API without restructuring.
BUILT FOR YOUR KIND OF AGENT
If it takes action,
it needs a QA loop.
For founders, CTOs, and engineers shipping agents into real workflows.
AI Support AgentsAI SDRsVoice AgentsWorkflow AgentsBrowser AgentsResearch AgentsCoding AgentsRAG SystemsMulti-Agent Systems
BEFORE YOU CONNECT
A few things
worth knowing.
An early product. A clear starting point.
Is this a self-serve product?
Not yet. We're building Tensile with early teams, starting with a hands-on, free agent QA audit. The product previews show the workflow we're working toward; the first step is testing your staging agent together.
What do you need to test our agent?
A staging or test API endpoint, dedicated test credentials, a request/response schema, and a description of expected behavior. We'll also need the constraints you care about, such as when a refund is allowed or what counts as a completed task.
How is this different from an eval dataset?
An eval dataset checks cases someone already wrote down. Our approach generates synthetic users and multi-turn scenarios around your agent's capabilities, explores edge cases, and repeats failures to measure how often they occur. It can complement your existing eval suite.
Do we need two agent versions?
Two versions let us compare a baseline and candidate on the same scenarios. If you only have one version, we can start by testing its behavior against your stated constraints and identifying reproducible failures.
Do you need production data or a new SDK?
The initial audit uses synthetic conversations and your existing staging API. It does not require production customer conversations or an SDK. Use dedicated test accounts, and keep any tool actions inside a safe test environment.
SHIP WITH FEWER SURPRISES
Your customers shouldn't
be your QA team.
Let synthetic users find the broken flow
before a real customer does.