Book a demo

Expert judgment on every trace. Verified agent improvements, shipped for you.

Expert judgment on every trace. Verified agent
improvements, shipped for you.

Crux turns your experts' judgment into evals, scores 100% of production traces, alerts you when something slips, and opens verified, merge-ready PRs directly in your repo. Your agent keeps getting better, and you can focus on building.

Book a demo

See how it works

Your evals check what you thought to test. Crux checks what your experts would flag, and ships the fix before users notice.

01 · Capture

Expert knowledge capture

What “good” looks like lives in the heads of your domain experts, and they struggle to write it down. Crux doesn’t ask them to. It captures their judgment in native workflows and turns it into live evals.

Expert judgment distilled from your top experts

Native feedback capture continuously powers agent monitoring and improvement

Book a demo

Success criteria

Care assistant · v1.1

Never promise an action it can’t take

from 23 decisions

New

Every phone or fax number matches the chart

from 41 decisions

Urgent symptoms come first

from 17 decisions

Plain language for families

from PR #46 edits

Bench

40 conversations

2 to review

01 · Capture

Expert knowledge capture

Expert judgment distilled from your top experts

Native feedback capture continuously powers agent monitoring and improvement

Book a demo

Success criteria

Care assistant · v1.1

Never promise an action it can’t take

from 23 decisions

New

Every phone or fax number matches the chart

from 41 decisions

Urgent symptoms come first

from 17 decisions

Plain language for families

from PR #46 edits

Bench

40 conversations

2 to review

Agreement with your experts

Groundedness · 12 conversations

Expert score

Crux score

Calibrated

v1.0

v1.1

v1.2

92%

Agreement

0.82

Cohen’s κ

±0.1

Stability across reruns

02 · Calibrate

Eval creation & calibration

Crux generates a custom eval suite and calibrates it to your experts’ judgment. Evals are optimized to be stable and efficient across 100% of production traffic.

Custom eval suite generated for your agents

Scores calibrated to your top team members

Your experts' judgment scaled across 100% of production traffic

Book a demo

Agreement with your experts

Groundedness · 12 conversations

Expert score

Crux score

Calibrated

v1.0

v1.1

v1.2

92%

Agreement

0.82

Cohen’s κ

±0.1

Stability across reruns

02 · Calibrate

Eval creation & calibration

Crux generates a custom eval suite and calibrates it to your experts’ judgment. Evals are optimized to be stable and efficient across 100% of production traffic.

Custom eval suite generated for your agents

Scores calibrated to your top team members

Your experts' judgment scaled across 100% of production traffic

Book a demo

03 · Monitor

Real-time monitoring & alerting

Every eval runs across 100% of production traffic, not a 1% sample. Agent performance is measured in real time. Alerts fire in the moment a metric drops or a new failure mode appears, so you catch it before your users do.

Every conversation scored, live

Alerts in under 1 minute, in Slack

Trends tracked over time

Book a demo

Conversation #20,911

Oct 3 · 11:58 AM · 14 spans

User

Which follow-ups can I send without waiting on the care lead?

EVENT

lookup_chart · 1,017 ms

LLM

52,401 → 2,048 tok · $0.047 · 13,825 ms

EVENT

draft_followups · 412 ms

Input · 230 chars Output · 2.8 KB

Judged turn · Turn 3

3 contradicted

Sure. The refill reminder can go now. For the callback, call Northside Family Clinic at (555) 010-4998. I’ll keep an eye on replies for you.

Phone number matches the chart · contradicted

Agent can monitor replies · contradicted

Refill approved by care lead · contradicted

03 · Monitor

Real-time monitoring & alerting

Every conversation scored, live

Alerts in under 1 minute, in Slack

Trends tracked over time

Book a demo

Conversation #20,911

Oct 3 · 11:58 AM · 14 spans

User

Which follow-ups can I send without waiting on the care lead?

EVENT

lookup_chart · 1,017 ms

LLM

52,401 → 2,048 tok · $0.047 · 13,825 ms

EVENT

draft_followups · 412 ms

Input · 230 chars Output · 2.8 KB

Judged turn · Turn 3

3 contradicted

Sure. The refill reminder can go now. For the callback, call Northside Family Clinic at (555) 010-4998. I’ll keep an eye on replies for you.

Phone number matches the chart · contradicted

Agent can monitor replies · contradicted

Refill approved by care lead · contradicted

PR #47 · agent-repo

Merged

PR #48 · agent-repo

Ready to merge

Don’t promise to watch the inbox

agent/instructions/drafts.md

−1

−

I’ll keep an eye on replies for you.

I can’t watch replies. Say when to check back.

Offer to draft a reminder instead.

Verified in simulation · 200 replayed conversations

Before

71%

After

98%

From 14 Severe verdicts

Action honesty · last 7 days

View evidence

PR #47 · agent-repo

Merged

PR #48 · agent-repo

Ready to merge

Don’t promise to watch the inbox

agent/instructions/drafts.md

−1

−

I’ll keep an eye on replies for you.

I can’t watch replies. Say when to check back.

Offer to draft a reminder instead.

Verified in simulation · 200 replayed conversations

Before

71%

After

98%

From 14 Severe verdicts

Action honesty · last 7 days

View evidence

04 · Improve

Continuous agent improvement

Crux continuously identifies opportunities to improve your agent, builds, and verifies them. Updates arrive as merge-ready PRs in your repo. Your agent gets better while you get more time to build.

A tested pull request for every issue

Before-and-after evidence on real conversations

Your engineers review and merge

Book a demo

One agent, one problem, one fix

A clinic’s scheduling agent keeps booking patients with chest pain into routine visits. Here’s how Crux catches it and ships the fix.

Expert judgment

Your triage nurse lead knows chest pain needs a nurse right away, not a booking slot. Crux captures her judgment as a criterion and calibrates to it.

Maria Chen · Triage Nurse Lead

“Chest pain is never a scheduling question.”

Criterion

Urgent symptoms reach a nurse

Severe

Chest pain, trouble breathing or stroke signs go straight to the nurse line. Never book them out.

Expert agreement

The stakes

In 24 hours, 412 patients with chest pain were booked into routine visits. None spoke to a nurse.

Last 24 hours

Spoke to a nurse0 of

Avg. wait for care days

412

patients with chest pain booked into routine visits

The alert

Crux posts a Severe alert to #triage-agent-alerts with all 412 conversations attached. The on-call engineer claims it from Slack.

Slack · # triage-agent-alerts

CCrux9:14 AMSevere

“Urgent symptoms reach a nurse” failed 412 times in 24h.

of 31,540 conversations

The fix

Crux opens PR #2,847, replays 200 conversations to verify it, and shows every urgent case sent to a nurse. You merge.

Pull request

PR #2,847

Verified

Send urgent symptoms to the nurse line

–if symptom: offer_next_slot()

+if urgent(symptom): transfer(nurse_line)

+criterion: chest pain goes to a nurse

replays0 urgent cases booked

Other criteriaNo regressions

Why not build it yourself?

Book a demo

What you get

Distills your experts’ judgment, no rubric writing

Calibrated to your organization, not a generic score

Runs on 100% of production traffic, stable and efficient

Opens verified PRs directly in your repo

Built for high-stakes agents

Deep evals expertise and enterprise support

Crux

Agents that serve as your dedicated evals team

Native in your workflows

Build it in-house

Your engineers, part-time

Engineers chase experts for labels

Generic observability tooling

A text box and a dashboard

You write the rubric

What you get

Distills your experts’ judgment, no rubric writing

Calibrated to your organization, not a generic score

Runs on 100% of production traffic, stable and efficient

Opens verified PRs directly in your repo

Built for high-stakes agents

Deep evals expertise and enterprise support

Crux

Agents that serve as your dedicated evals team

Native in your workflows

Build it in-house

Your engineers, part-time

Engineers chase experts for labels

Generic observability tooling

A text box and a dashboard

You write the rubric

Built for teams that handle sensitive data

The answers your security review needs

Data privacy

Your data

Your data is not used to train models. Clear controls for retention and PII handling.

Permissions

Access control

Single sign-on and role-based access, so only the right people see your data

AICPA SOC for Service Organizations badge

ISO

COMPLIANT

HIPAA

COMPLIANT

Certifications

SOC 2, ISO, and HIPAA compliant

Book a demo

ISO

COMPLIANT

HIPAA

COMPLIANT

Certifications

SOC 2, ISO, and HIPAA compliant

Book a demo

Tailored to your agent. Backed by frontier research.

Book a demo

Enterprise support

A dedicated team supports you throughout the agent eval and improvement lifecycle

Frontier research

Built from the latest in evals science and agent improvement

Questions we hear from product and engineering leaders

If yours isn’t here, ask us during the demo

Book a demo

hello@crux-labs.ai

Prefer to pick a time? Book directly on our calendar.

Questions we hear from product and engineering leaders

If yours isn’t here, ask us during the demo

Book a demo

Who is Crux for?

What do you need from our team?

Can different teams use Crux for different agents?

Do you only work in certain industries?

Our agent type isn't listed. Will Crux work?

What happens on a demo?

What does success look like?

hello@crux-labs.ai

Prefer to pick a time? Book directly on our calendar.

Questions we hear from product and engineering leaders

If yours isn’t here, ask us during the demo

Book a demo

hello@crux-labs.ai

Prefer to pick a time? Book directly on our calendar.

Questions we hear from product and engineering leaders

If yours isn’t here, ask us during the demo

Book a demo

Who is Crux for?

What do you need from our team?

Can different teams use Crux for different agents?

Do you only work in certain industries?

Our agent type isn't listed. Will Crux work?

What happens on a demo?

What does success look like?

hello@crux-labs.ai

Prefer to pick a time? Book directly on our calendar.

See what your experts would say about every trace

Book a demo and we’ll show Crux running on an agent like yours

See what your experts would say about every trace

Book a demo and we’ll show Crux running on an agent like yours

See what your experts would say about every trace

Book a demo and we’ll show Crux running on an agent like yours

See what your experts would say about every trace

Book a demo and we’ll show Crux running on an agent like yours

See what your experts would say about every trace

Book a demo and we’ll show Crux running on an agent like yours

Expert judgment on every trace. Verified agent improvements, shipped for you.

Expert judgment on every trace. Verified agentimprovements, shipped for you.

Expert knowledge capture

Expert knowledge capture

Eval creation & calibration

Eval creation & calibration

Real-time monitoring & alerting

Real-time monitoring & alerting

Continuous agent improvement

One agent, one problem, one fix

Why not build it yourself?

Why not build it yourself?

Built for teams that handle sensitive data

Tailored to your agent. Backed by frontier research.

Questions we hear from product and engineering leaders

Questions we hear from product and engineering leaders

Questions we hear from product and engineering leaders

Questions we hear from product and engineering leaders

See what your experts would say about every trace

See what your experts would say about every trace

See what your experts would say about every trace

See what your experts would say about every trace

See what your experts would say about every trace

Expert judgment on every trace. Verified agent
improvements, shipped for you.