Benchmark
Solve the escalations that bounce between teams for days
Watch BuildWright trace a real customer-reported bug to the exact file and method the shipped fix changed, using nothing but the customer’s description.

The problem · ERPNext
“The invoice totals never reconcile…”
Off by a few cents, only sometimes, and nothing in the ticket says where to look. This kind of ticket bounces between teams for days.
“On some multi-item invoices, the item-wise tax breakup doesn’t add up to the total tax. It’s off by a few cents, only on certain rate and quantity combinations. The header total looks right; the per-item breakdown doesn’t reconcile.”
The solution
BuildWright investigated
BuildWright traced the missing cents to lines 540 to 556 of the set_item_wise_tax() method in taxes_and_totals.py.
Tax was rounded on each line separately, so small differences drifted from the header total. ERPNext pull request #53925 changed the lines BuildWright named.
- File
- erpnext/
controllers/ taxes_and_totals.py - Method
- set_item_wise_tax()
- Lines
- 540–556
- Cause
- Tax rounded on each line separately. Small differences add up and drift from the header total.
Matched the shipped fix. ERPNext PR #53925 changed the same lines.
Technical detail
It works the way your team would
A read-only index of the code, related knowledge, and a grade on the evidence. The only input was the customer’s description.
With its knowledge layer switched off, BuildWright’s production agent matches the reference agent on τ²-bench, Sierra Research’s public benchmark for AI support agents.
| τ²-bench domain | BuildWright | Reference agent |
|---|---|---|
| Retail | 0.82 | 0.81–0.83 |
| Airline | 0.79 | 0.80 |
Resolution rate across 50 tasks per domain, graded automatically.
Results
BuildWright doesn’t mark its own homework
Not enough evidence to name a root cause
no indexed code repository
Connect the repository for this service, or add the error log from the time of the report, and I will investigate again.
A tool that says “I don't know”
In one run it reported “no indexed code repository” instead of inventing a file path.
BuildWright used a past resolution
Same rounding issue, different customer
Learned resolutions surface the facts it needs roughly 3x more reliably the next time.
Faster with every fix
Once it learns a resolution, the facts it needs surface roughly 3x more reliably the next time.
- ShadowDrafts only. Nothing posts.
- CopilotA person approves every post.
- AutopilotOnly for proven case types.
Read-only by design
It never changes code, data or production. Your team sets the trust level.

Get started
Grade BuildWright on your own escalations
Read-only access, personal data masked, running in your own cloud. Start on the sample app today, or bring a ticket your team already solved and judge what it finds.
Not ready for a call? Send us a note
BuildWright could not support a diagnosis