Inferred workflow

We watched 142 traces from your agent. Here's what we think it does — review and edit before we run baseline.

1 Connect source

2 Inferred workflow

3 Generated eval set

✓

Inferred from 142 traces · spanning 47 hours · 89% confidence

Your agent appears to be an order intake assistant handling new customer orders coming through chat and email. We detected 3 distinct steps, 6 tool calls, and 1 LLM model. Edit anything below — these are our best guesses, you know your agent.

Inferred workflow definition

Edit fields directly

Name

Order intake assistant edit

What it does

Handles inbound customer orders from chat and email. Looks up SKU availability and pricing, validates payment terms against the customer record, and drafts an order confirmation. Routes to a human when payment terms differ from contract.

Outcome metric (auto-detected)

Order acceptance rate · 81.3% over last 142 traces
Inferred from order.status = "accepted" events. Pick a different metric →

Inferred steps · 3

Parse customer request step.parse

Extracts SKUs, quantities, customer ID from free-form input

142 / 142 always

Resolve customer + SKUs step.resolve

Calls inventory and CRM tools in parallel

142 / 142 always

Draft confirmation or escalate step.respond

Branches: 87% accept-and-confirm, 13% escalate to human

142 / 142 branching

+ Add step manually

Failure signals to watch (auto-suggested)

order.escalated_to_human order.payment_terms_mismatch tool.inventory_lookup.timeout customer.followup_within_24h + add signal

Detected tools · 6

inventory_lookup 412 calls · avg 218ms

customer_record_get 142 calls · avg 89ms

price_calculator 142 calls · avg 41ms

order_create 123 calls · avg 312ms

human_handoff 19 calls · avg 4ms

email_send 119 calls · avg 184ms

Detected models · 1

claude-sonnet-4-6 142 / 142 traces

Suggested view

ConversationView

Best fit: chat-style intake with branching outcomes

88% confidence

We picked this from your trace shape. Choose a different one →

Next: what we'll do

Generate an eval set from your traces (expected: ~30 cases)
Run a baseline pass on those cases — establish the floor
Watch new traces for failure clusters
Propose a first improvement once we have signal