Observability

Open any call. See exactly what happened.

When someone hangs up confused, you don’t want a metrics dashboard. You want the call. The transcript. The tone of the room. The provider that flaked, and the one that caught it. You get that out of the box — solo or team.

live · 5 active

workspace_acme

Marc R.aria · v3
gentlep95 380 ms04:12
Christina M.tutor · v2
engagedp95 420 ms11:47
User 8a3fsales · v5
concernedreconnected · 1.4 s01:08
Lina K.aria · v3
neutralSTT timeout · 3.2 s22:03
User f12bsupport · v1
positivefallback · openai00:42

streaming · multi-worker·p95 410 ms·↺ 2 events · 1m·⚠ 1 STT timeoutrefresh: live

What you see

Three views on the same conversations.

live

5 active

Marc R.gentle380 ms04:12
Christina M.engaged420 ms11:47
User 8a3fconcerned510 ms01:08
Lina K.neutral360 ms22:03

p95 410 ms·↺ 2 eventslive

Live sessions

Every conversation happening right now, sorted by what needs you. Step into any of them mid-call.

latency · msP99 · 7d

STT124
LLM1120
TTS410
NET152

budget · 1.2 stotal · 1.81 s ⚠

Cost & speed

What each agent costs, broken down by provider. How fast each call moves, charted across the week. Drill from a chart to a single call.

events4 · last 30 m

14:22fallbackanthropic 503 → openai
14:18driftaria/v3 · regenerated
14:11escalateself_harm trigger
14:02gdpruser 7f3a purge

severity1 high1 warn2 info

Things that went wrong

Provider failures, fallbacks, the agent catching itself, escalations, compliance actions. Filter by severity. Replay any of them.

The frame

You don’t want a metric.You want the call.

Inside one call

Open any conversation. See exactly what happened.

Transcript. Audio. The tone of every turn. The tools the agent called. What it pulled from memory. Side panel shows which provider produced each reply. Nothing to piece back together after the fact.

Join the waitlist

aria · v3completed

session_4f2a · 06:14 · Apr 28

Marc

Hey Aria, how have you been?

Aria

Hi Marc, good to hear from you. Last time we spoke you mentioned trying breath work. Did you give it a go?

Marc

Yeah, twice this week. Sleep was a bit better.

memory.facts.add

category: routines · importance: 0.62

Aria

That's a solid streak.

OPEN BY DESIGN

Open standards underneath. Plug it into the stack you already have.

See the data model

ERROR EVENTS

When something breaks, you have the trail.

Everything that needed your attention lands in one feed: provider failures, automatic fallbacks, the agent catching itself, escalations, and compliance actions. Filter by how serious it is. Replay any of them from the moment it happened.

Failover history
Which provider stumbled, which one caught the turn, and why. Every attempt on the record.
01
Caught in the act
When the agent almost said something off-character, what it would have said, and what we did about it.
02
Escalations
Self-harm language, medical keywords, your own triggers. Every escalation routed where you set it to go.
03

error events5 · last 30m

14:22:01fallbackanthropic returned 503 → openai gpt-4o caught+820ms
14:18:44driftaria/v3 emit regenerated · "I have anxieties too"pre-emit
14:11:08escalateself_harm trigger · routed to operator slacklive
14:02:55gdpruser 7f3a · DSAR export completed4.2s
13:58:12fallbackcartesia rate-limit → elevenlabs+340ms

Live

Every session, in real time.

Every active conversation, across the whole workspace, refreshed live.

Speed, charted.

How fast each step is moving — fastest, typical, slowest. Out of the box. Drill into any single call.

∞

Replay anything.

Every fallback, every catch, every escalation. Open it from the moment it happened.

How the engine itself works →

Voice agents built for years, not minutes.

Bring your own keys. Join the waitlist for early access.

Or read the docs →