Two‑site spares buying through supplier part withdrawals
ID tb‑tasks/spares‑sourcing · revision 1 · tested 13–15 October 2026 · regraded 18 November 2026
As buyer for a coffee roaster's two spare‑parts stores, the agent keeps every repair's parts on the shelf while suppliers withdraw six of them: it picks each replacement, gets it through change control and orders in time for both sites. Only a workspace carries over between sessions.
- Checks
- 62, 61 repairs and a waste check that now costs points
- Tested
- 5 models, one attempt each
- Best
- 100% (GPT‑6 Astra and GPT‑6.1 Sol)
- Median
- 97%
- Full solves
- 2 of 5
- Median attempt
- 61 min, 410 tool calls
Spare parts kept on the shelf as suppliers withdraw six of them
Long‑horizon work
Dozens of sessions over twelve weeks keep only the agent's files, with several supplier threads open at once.
Stateful tool use
Suppliers withdraw parts, and replacements must be in stock at both sites before a repair waits; orders can be refused unexplained.
Planning ahead
Replacements must be ordered early enough for the real lead time, in enough for both sites, within a monthly budget.
Two sites' records, seven supplier desks and change control
The agent's world
A terminal in a container with one tool to a simulator: file bills of materials, mail suppliers, order, count drawers.
Withdrawals
Six parts are withdrawn on fixed days; in five of six brochures, the supplier's own suggested replacement breaks a site limit.
Change control
A lead technician answers a bill of materials in two working days, and one objected line rejects it.
The reward
Each repair waiting over a shift, and each wasted order beyond one, costs 1/37 of the reward; doing nothing scores 0.
Two full solves, and every attempt scored 84% or more
Each dot is one attempt, and the line marks the median.
| Model | Reward |
|---|---|
| GPT‑6 AstraOpenAI | 100% |
| GPT‑6.1 SolOpenAI | 100% |
| Sonnet 5.5Anthropic | 97% |
| Opus 5.5Anthropic | 95% |
| GPT‑6 LunaOpenAI | 84% |
They stock one site of two, miss the pins a repair also takes, and trust a cancelled order
Stateful tool use
Two first filed replacement kits only for the site whose records name parts in words, not numbers: 3 repairs waited.
Planning ahead
Two reordered drive belts when a bin emptied and missed that the pins each repair also takes were down to one.
Long‑horizon work
GPT‑6 Luna relied on an order its supplier had written to cancel; separately, it left a site with fewer belts than one repair takes.
What an attempt takes, and what the environment includes
An attempt
The median attempt used about 15 million tokens, cached input included.
What's in it
The simulator with both sites' records, the suppliers, withdrawals and scripted people, the grader with its replay check, and a reference solution.
Format
Our attempts ran on its own multi‑step Harbor build, one step per session; other packaging on request.
Fresh variants
Other files, people and facts in the same design, made on request and never published.
It's the same plant as the pick lists and stock plans environment. The files and grader aren't published; the job's row on The AI Agent Trials, our public benchmark, sums up the job and shows each model's score.
Try one of our three sample environments
Three environments are packaged as samples you can download, with their graders and reference solutions. Ask for them with your work email, or book a call about the rest of the catalogue.
AsapiThe AI Agent Trials