Built By
Deploy your agents, without the firefighting.
Radiant observes every action your agents take across phone, browser and text, detects bugs and frustrated users, and tests the fix with simulated users.
Trusted By
Backed By
Agents used to answer questions. Now they complete tasks.
They take the call, fill in the portal, chase the reply and book the appointment.
One job can run for days, across phone, browser, text and email.
When it fails, the customer notices first.
The call went fine. The job didn't.
Every check passed. The job still failed.
The call scored well and the transcript reads perfectly. The agent promised a follow-up by Thursday. The follow-up never happened, and you heard about it a week later from the customer. Failures hide between steps and channels, and surface days later.
Debugging means reading transcripts across five tools.
The voice log is in one tool, the traces in another, the workflow history in a third. Someone lines them up by timestamp to work out what happened. Nobody sees the whole job.
You can't put a phone call in CI.
Testing a journey that crosses a phone call and a web form means someone dialling in and someone clicking through, every time. It takes an afternoon, so it stops happening.
Every change is a gamble.
A prompt edit fixes the date bug and breaks the address step. A model upgrade is faster and quietly skips a tool call. So teams freeze the agent and stop shipping.
Your best engineers became babysitters.
They replay calls, watch dashboards and patch prompts late at night. You hired them to build agents. Now they mind them.
Observe. Simulate. Improve.
- 01 Observe Trace every job and see the step that broke.
- 02 Simulate Rehearse every change with simulated customers.
- 03 Improve Fix the cause, then prove it on the same journey.
01
Observe. See the whole job.
Every call, tool call, browser action and workflow step, joined into one view across days.
Referral 4821
Not booked Intake was rejected for a missing email, so the booking call never ran.
More to the right · drag to pan
02
Simulate. Rehearse every change.
Simulated customers call, click and wait, and every run is judged on the outcome.
Simulator · Patient Called in for 5m 18s
Referral intake agent
✕ submit_intake rejected
submit_intake rejected after the callSimulator · Operator
Opens referral queue in web app: nothing new
Opened the referral queue: nothing newSimulator · Office
✕ Booking call never arrived
Waited 19m. The booking call never arrived.ExpectedPatient booked
GotNo booking call. Stopped at the 25 minute limit.
03
Improve. Find the cause. Prove the fix.
Get an answer that cites the traces, hand the fix to your coding agent, and re-run the journey.
Why are referral bookings stalling since Tuesday?
submit_intake was rejected in 6 of 10 attempts1; 4 had no retry2.
Likely cause: conflicting retry instructions in the prompt3. Line 18 says never ask the caller to repeat details
. Line 31 says retry failed submissions with corrected fields
. With no email on file, the agent can
do neither, so it stops.
- 110 jobs since Tuesday · submit_intake · 6 × 422 "email is required"
- 24 of the 6 failed jobs · no retry span after the rejection
- 3Prompt v42 · line 18 vs line 31
Hand fix to your coding agent Re-run the same journey
Debug and simulate agents that work in the real world.
Voice, browser and text
Radiant traces every call, click and message in one job, so you debug the whole journey, not one channel.
Multi-day, multi-party
Replay and simulate jobs that wait on customers, staff and other companies for days.
Integrate with one line
Already emitting OpenTelemetry? Point the exporter at Radiant.
OTEL_EXPORTER_OTLP_ENDPOINT=https://radiant.your-cloud.internal:4318Show full setup
# Standard OpenTelemetry exporter settings
OTEL_EXPORTER_OTLP_TRACES_ENDPOINT=https://radiant.your-cloud.internal:4318/v1/traces
OTEL_EXPORTER_OTLP_TRACES_PROTOCOL=http/protobuf
OTEL_SERVICE_NAME=claims-follow-up-agent
# Python: the standard SDK reads the settings above
from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
provider = TracerProvider()
provider.add_span_processor(BatchSpanProcessor(OTLPSpanExporter()))
trace.set_tracer_provider(provider)Runs in your cloud, if you want
Use Radiant's hosted platform, or self-host it in your own cloud so traces, recordings and transcripts never leave your environment.
Works with your stack.
Bring the agents you already run. Radiant reads standard OpenTelemetry traces, so there is nothing new to adopt.
- Vapi
- Retell
- LiveKit
- OpenAI
- Amazon Bedrock
- Your own agents
OpenTelemetrystandard traces
Radianthosted or self-hostedComing soon
RL environments, built from your production data.
Radiant will turn real jobs and simulations into reinforcement-learning environments, so you can train and evaluate your agents on the work they actually do.
Questions teams ask first.
What do we need to connect?
OpenTelemetry traces from your agents. If you already emit them, point your exporter at Radiant. If you do not, we will show you where to add them.
Where does our data live?
Radiant is hosted by default, so there is nothing to run. If you need it, you can self-host Radiant in your own cloud instead, and traces, recordings and transcripts never leave your environment.
Which stacks do you support?
Voice platforms such as Vapi, Retell and LiveKit, models on OpenAI and Bedrock, and custom agents you built yourself. If it emits OpenTelemetry traces, Radiant can follow it.
How do simulations stay safe?
Simulations use simulated customers, not real ones. They call your test numbers and work in the environments you choose. No real customer is contacted.
How much does it cost?
Talk to us. Tell us about the job and we will give you a straight answer on the first call.
We build in the open.
We're big believers in open source. Our agent harness is already open, and the core of Radiant is next. Read the code, don't take it on faith.
Polaris
Our agent harness, written in Rust. Agents are explicit graphs you can read, sessions checkpoint and roll back, tools run with scoped permissions, and traces export to OpenTelemetry.
Explore PolarisHubble
The core of the Radiant platform: the part that captures and stitches together every job your agents run. We're open-sourcing it soon.
Get notified when Hubble is open-sourcedReady to stop firefighting your agents?
We're opening up our beta to more users, reach out to join the waitlist!
Or book a demo