01 · Test
Test on simulated users
A model plays the person, another your tools; a judge grades.
Flow cloud
Host it. Test it. Watch it.
Simulated people test your agent; every real conversation gets graded.
What Flow cloud does
Three jobs. One dashboard.
02 · Watch
Watch and grade every conversation
Every reply graded good, wrong, unhelpful or unsafe.
03 · Host
Host it
Run your agent beside its tests and dashboard, or self-host.
Everything in it
The details.
- Scenario filesOne YAML file per situation: person, goal, world, expectations.
- Every channelEach scenario runs once per channel, with its own result.
- Simulated toolsA model plays your tools, slow answers and errors included.
- A judge with reasonsPass or fail per expectation, with one sentence why.
- LatencyFirst-reply time per message: average, p95, min, max.
- Cost per runAgent and test model costs, shown separately.
- ReplayWatch any run again, message by message.
- Laptop or CIA failed test exits non-zero, so CI stops the build.
- Traces per turnModel calls, tool calls, errors and retries inside each reply.
- IssuesFailures grouped, ranked by conversations touched, with a suggested fix.
- VersionsWhat runs in staging and production, and how each did.
- Any stackSend one OpenTelemetry trace per turn; it is graded the same.
Security
Your customers’ words, kept apart.
- Message text encrypted, a key per account
- Every view of a conversation logged
- Row-level security between accounts
- Separate trace, CI and messaging keys
- Test data only until production is on
Know your agent works before customers do.
Host, test and watch your agent in one place. Sign in with your Flow login.