Trace

See every session.

Every run, every check, every human step, with the version that produced it.

Phinite TraceTriage v4.3 live
Sessions · today0 checked
#48211Charged twice, order 4821Verified
#48212Priority with two SLAsCorrected
#48213Refund $260, over limitPerson
#48214Change billing addressVerified
#48215Where is my order?Verified
Replay · last 7 days · 6,120 casesVerified autonomous
v4.2 · live0%
v4.3 · candidate0%
Drift on category step-0.8 pts
Later corrections, 14 days11
No vetoNo regressionBest configuration
Bridge metric · tickets handled correctly, no person
Verified autonomouswhole cases, all five conditions
0%
Critical misses0%
Wrong actions0%
SLA breaches0
Guardrails shown0 sessions

Phinite Trace is the proof. Every decision on a trace, every version tested before it ships, and an alert when a live agent drifts from what you approved.

What Phinite Trace gives you

For quality, risk and platform teams.

route
retrieve
tool call
policy
resolve

Session traces

Replay every run, check and human step, with the version behind it.

2 beforedeploy3 after

Pre and post evals

Two gates before release. Three checks after it.

Veto: noneRegression: noneOptimize: config B wins

Veto, Regression, Optimize

No veto, no regression, and the best configuration wins.

Sustained

Drift alerts

Know when a live agent stops behaving like the one you approved.

Measuring success

“It ran without errors” is not success.

An agent can finish cleanly and still make the wrong call. Trace measures whole cases, from intake to confirmed action.

A case counts once

However many runs it takes. Step accuracy is not a business result.

Five conditions, all true

A case is verified only when every one holds.

In scopeAction confirmedCritical checks passedNo person stepped inNo later correction

Guardrails beside every number

Autonomy is always shown with critical misses and wrong actions.

1,000 in-scope tickets0
750 right on the first pass200 fixed by a correction agent50 sent to a person
0%First-pass rate: what one agent did
0%Verified autonomous: what the business got

verified autonomous rate = verified autonomous cases ÷ in-scope cases

Before and after

Proven before release. And after.

Two gates before deployment. Three monitoring stages after it.

Before deploymentAfter deployment
01

LLM playground

Best configuration, with evidence.

02

Test simulation

Thresholds on real scenarios.

Deploy
03

Operational signals

Every session checked.

04

Drift detection

Sustained change, not noise.

05

Live eval metrics

Cause found, fix recommended.

0%of sessions get signals and drift checks
0%of high-risk cases get deep evals
0fixes reach production without re-testing
Three gates

Nothing ships on a hunch.

Every version passes all three before it is promoted.

Veto
Critical checks100%
Policy breaches0
v4.1 · wrong refund actionvetoed

Veto

Nothing promotes with a veto.

Regression · live vs candidate
Scope
Action
Autonomous

Regression

Nothing gets worse than what is live.

Optimize · three configurations
Config A93.8
Config B95.6
Config C94.4

Optimize

The best configuration wins, on evidence.

Inside Phinite Trace

Four views of agent quality.

Test. Compare. Monitor. Your customers only meet your best build.

Scenario suites

Run the candidate on real scenarios before anything ships.

Suite · Ticket triage · 240 scenariosv4.3 candidate
In scope correctly identified99.2%
Required action confirmed97.5%
Critical checks passed100%
Wrong actions0.4%

Session replay

Replay last week's cases on both builds, side by side.

Replay · last 7 days · 6,120 casesv4.2 vs v4.3
v4.2 (live) verified autonomous94.1%
v4.3 (candidate) verified autonomous95.6%
Drift on category step-0.8 pts
Later corrections (14-day window)11

Verified autonomous rate

One bridge metric, always shown with its guardrails.

Bridge metric · handled correctly, no personGuardrails shown
Verified autonomous95%
Critical misses0.1%
Wrong actions0.4%
SLA breaches0

Drift alerts

Sustained change on a live agent, with the fix and the next step.

Alerts · Triage v4.2 · live3 open
Priority mis-set when two SLAs applySustained 3 days
Routing to retired queueFix re-tested · passed
Out-of-scope language detected lateScope rule proposed
Next stepPromote v4.3
One platform

Build, run, govern and improve agents on one platform

Agent Graph

Design agents and hand-offs on one canvas.

View Agent Graph
Agent Graph Studio canvas

Agent Runtime

Run every version on your own cloud.

View Agent Runtime
Runtime · refund-agent
Devv14testing
UATv13signed off
Productionv13running
1,240runs today · 99.9% up

Agent Trace

Follow every step, tool call and decision.

View Agent Trace
Trace #4821 · 2.21s · $0.004
ticket received
triage agent
get_order
policy check
issue_refund
reply sent

Phinite Aura

Describe an agent and get the graph.

View Phinite Aura
Phinite AuraAgent Graph Context Active
An agent that sends satisfaction surveys and follows up on low scores.
Great idea. Here is what I understand:
  • Sends surveys on a schedule
  • Finds low scores and trends
  • Follows up and updates the CRM

Eval Studio

Test every version before it ships.

View Eval Studio
Eval Studio

Dev Studio

Write and publish your own tools.

View Dev Studio
email_tool.py · v11def main(inputs, env):2  to = inputs.get("recipient")3  if not to:4    raise ValueError()5  return send(to)
Publish tool
Current V1New V2
Publish v2
FAQ

Everything about Phinite Trace.

The questions teams ask us most. Anything else, ask us in a demo.

What is Phinite Trace?

The record and the test of everything an agent did, before and after it ships.

What is the verified autonomous rate?

Verified autonomous cases divided by in-scope cases. A case counts only when all five conditions hold.

What happens before a version ships?

It runs in the LLM playground and a test simulation, then must pass three gates: no veto, no regression, best configuration.

How does drift detection work?

Every live session is checked against the version you approved. Sustained change raises an alert, not noise.

Can a fix go straight to production?

No. Every fix is re-tested before it reaches production.

Prove what your agents changed in the business.