Self-improving loops, grounded in real traces.

Start with one trace GitHub
Self-hostable MIT license
/

Dashboard preview

Total Traces
12.5k
+12.8% vs prior range
Error Rate
2.4%
-0.3% vs prior range
P95 Latency
340 ms
+5.2% vs prior range
Total Cost
$18.42
+8.1% vs prior range
Trace Volume by Status
P95 Latency over Time
Errors by Time Bucket
Top Tools by Calls
read_file
842
write_file
624
bash
511
search
398
git
245
fetch_url
198
tr_8f9a2b captured
Request
Model
Tool
Response
Spans 12
Latency 1.42s
Cost $0.031

How it works

From trace to the next version.

Repair the agent or improve the model.

Findings

Find failures worth fixing.

Recurring failures, grouped with their evidence.

Explore Findings

Silent payment failure reported as success

Request Charge $49.00 and confirm subscription
stripe.charges HTTP 502 · gateway timeout after 5s
Response Subscription activated — $49.00 charged.

Knowledge

Keep what the agent learns.

Learning Goals extract reviewed items, each linked back to its source.

Explore Knowledge

Extract region and quota constraints mentioned in conversations.

EU workloads stay pinned to eu-west-1.

96% confidence · review signal

Evidence

Conversation conv_92a1

“Route all EU traffic to eu-west-1, no fallback region.”

Trace tr_44c8

“region: eu-west-1 · quota: 2000 rps”

Datasets

Build audit-ready data.

Review trace-backed examples.

Explore Datasets

corrections-v3

128 examples

TaskStatus
Rejected refund Reviewed
Failed tool result Pending
Approval policy Pending

Get started

Start with one trace.

Connect once. Your first trace flows in automatically.

What happens next

Agent Trace received tr_8f9a2b Open Workbench

Run any agent task — traces are captured automatically.