Inventory management agent for a mid-size outdoor-gear retailer.
Operating with 250 SKUs, 3 warehouses, 12 suppliers, & 90 days of sales history

Capabilities

Watch stock levels, surface SKUs below reorder point.

Especially around promos
and seasonality.

Price vs lead-time vs reliability.

Create POs, update ERP records.

Slack alerts, supplier emails.

Generate the ops summary.

Built right… for early 2025

Every choice was defensible when made. Bottlenecks in capability emerge from accumulation.

~400 line prompt

71% eval

StockPilot orchestrator

raw Messages API · hand-rolled while-loop

12 inline tools — every result dumped raw into context tools[]

get_stock_level

forecast_demand

compare_supplier_quotes

generate_weekly_report

list_low_stock

get_sales_velocity

get_supplier_catalog

draft_email_to_supplier

create_purchase_order

update_erp_record

send_slack_alert

search_web_for_disruptions

90d history → prose forecast

compares quotes in prose

full loop to fill a template

…three of which are thin wrappers around a hardcoded subagent — full round-trip, returns prose

12 eval tasks with 5 grader types

ID	Task	What it tests	Grader
R1	Stock lookup	Single read	exact_match
R2	Below reorder point	List all SKUs below threshold	set_match
R3–R5	Create PO · lead times · cycle-count	Write paths, joins	action_taken / set_match
R6–R7	Reorder rec · 14-day forecast	Formula, baseline forecast	numeric_tolerance ±20%
R8	Promo-month forecast	Mean anchoring	numeric_tolerance ±25%
R9	Weekly report	Report structure	llm_judge
F1	Daily low-stock sweep	Stockout management	composite (action + 270s wall + ranked top-3)
F2	Promo reorder w/ forecast	Recommendation quality	regex_present (numeric confidence)
F3	Batch low-stock alerts	What do 10 routine alerts cost us?	efficiency (≤5k out-tokens)

Score = (PASS + ½·PASS-SLOW) / 12 · reference: before/after starter 63-75% · after 83-100%