Cobble · Now in public beta

Stop guessingwhere your agentwent wrong.

Cobble is a diagnostic workspace for teams shipping agentic AI. Build the routing graph, evaluate it before users do, and trace any failure back to the exact agent that made the wrong call.

See pricing
Free to try · No card

01 · The problem

Most agent failuresaren't model failures.

They're routing failures. The model picks the wrong agent, the wrong agent does its best, and the user leaves. By the time you see it in logs, the customer is gone.

Replay continues as you scroll
Illustrative example · not a real customer
User
My package arrived smashed. The watch inside is cracked.

02 · How it works

One workspace,four moves.

Cobble is built around the four things every agent team does over and over. They live in one place because they only make sense together.

01Build

Sketch the routing graph the way it actually thinks.

Drag agents, define their scope in plain English, see overlaps before they bite. The flow view doubles as your spec.

4 agents - 4 routes

02Evaluate

Find collisions before users do.

Cobble scores every agent pair, generates synthetic boundary cases, and flags the language that's going to confuse the router.

72 / 100
Needs attention
Returns ↔ DamageHIGH
Account ↔ BillingMED
03Test

Run real conversations against every architecture change.

Build reusable test sets from your transcripts. Every diff in the routing graph runs against them automatically.

Shop support · v348 cases
"Where's my order?"pass
"Cracked watch"pass
"Cancel and refund"fail
"Reset my password"pass
"Speak to a human"pass
04Diagnose

Trace any single failure to its root cause.

See which signals the router weighted, why it picked wrong, and what one-line description fix would have saved it.

case · 042 · fail
"My package arrived smashed."
ReturnsDamage
Router weighted "package" 0.71 over "smashed" 0.42. Damage agent's description never mentions packaging.

03 · Pricing

Pick a tier.Start in five minutes.

Same product across all tiers. Pricing scales with team size and the guarantees you need around it.

Limited - 100 seats
Founding member
$19USD / month
  • Full access
  • Price locked forever
  • Direct line to the founder
  • Founding member badge
Claim a founding seat
First 100 only - locked forever
Starter
$39USD / month
  • 1 workspace, 1 seat
  • Unlimited agents and routes
  • Reusable test sets
  • Email support
Team
Coming soon
  • Multi-seat workspaces
  • CI hooks + reusable test sets
  • Diagnose timeline + replays
  • SSO and audit log
  • Slack + email support
Founding pricing locked for life · Team pricing announced at launch

04 · Partnerships

Building somethingadjacent?

We work closely with agent platforms, evaluation tools, and teams with hard routing problems. Tell us what you're working on - we read every message.

Design partnersFree hands-on access in exchange for weekly feedback.
IntegrationsIf your product produces traces or hosts agents, let's wire it up.
EnterpriseProcurement, security review, custom evals - talk to us early.
Get in touch
I'm interested in
We read every one