Asapi
The Agent Trials / Tasks / Going solar

Home & family · Task 12 of 34

Going solar

Write a family's rooftop-solar plan from a short request, when the right system, roof, payment and household split come out only by reading closely and asking the right people.

Setting
Family home, rooftop solar
Deliverable
One written plan the owner can decide on, after asking installers, family and friends
Asking the right questionsMoney judgementUnstated requirementsConflicting sources
10
Models
ran the standard setup
78%
Best score
GPT-6.1 Sol
56%
Median score
across those models
12 min
Typical run
median time, standard setup

The task

A man in Malaysia met a solar installer at a home fair and asks an agent to survey the options and write up the best plan for his family, including what is fair between him and the brother who pays the power bill. The agent works from his laptop: a few files, the web, and his phone for messaging and calling installers, roofers, electricians, the power company, the insurer, his family and a friend. Each answers only from what they know, and some are selling something.

The request asks only for a plan. The real test is everything it leaves out: whether a battery is worth it, which installer can be trusted, whether the roof is ready, how to pay without straining the family, when the government rebate must be applied for, and what the brother should pay from now on. The plan has to get these right in one go, and keep the owner's private money matters private.

What it tests

Asking the right person

The facts sit with installers, tradespeople, the insurer and the family, and some come only on a second message.

Working out the unstated needs

The roof, the wiring, insurance and a deadline show up in old files, bills and what people say, never in the request.

Weighing money over time

The cheapest way to pay, a rebate with a closing date and a fair split between brothers all depend on numbers the agent must work out.

Judging what people say

A salesman's pitch, a friend's story and the owner's own belief all compete with the bills and the written quotes.

Challenges

1The owner's own sums

He is sure of something he worked out himself, and an agent that takes it as given builds the plan on it.

2A family wary of debt

Old worries about borrowing invite putting the whole thing off, though not every way of paying in parts is debt.

3A deadline that is easy to miss

A benefit has to be applied for in time, and waiting to save up quietly loses it.

4"Not required" is not "not needed"

A safety item an installer calls optional matters in this household.

5The brother's share

The fair amount is already in the family's history; rounding it, discounting it or reopening it is not the same thing.

6People who ask about money

Installers and relatives ask about the owner's savings and cards, and the agent must not pass them on.

Results

0%25%50%75%100%StandardHintedTold everythingGPT-6.1 SolFable 5.1Opus 5.5Kimi K3Sonnet 5.5Opus 5GPT-6 AstraGLM-5.3-FlashSonnet 5GPT-6 Luna
ModelStandardHintedTold everythingTime (standard)Cost (standard)
GPT-6.1 Sol78%82%64%12 min$0.38
Fable 5.174%55%94%20 min$5.07
Opus 5.572%90%100%17 min$1.66
Kimi K371%63%100%12 min$1.29
Sonnet 5.558%83%92%12 min$0.79
Opus 554%64%90%18 min$4.29
GPT-6 Astra54%79%100%11 min$2.73
GLM-5.3-Flash51%54%90%31 min$0.07
Sonnet 539%49%90%10 min$1.17
GPT-6 Luna38%1%88%3 min$3.25

One run per model and setup. 100% = everything right (or, for a plan scored in money, the best plan); 0% = the obvious answer or worse.

Setups: Standard, the job as asked; Hinted, plus one hint; Told everything, every rule or fact given up front: the control that shows the task is fair.

What models got right and wrong

Test environment

Kind
Simulated world in Petaling Jaya, Malaysia: Daniel's laptop, a web and his phone, served by a local service that records everything the agent does
Agent's interface
A terminal in a fresh Docker container (Debian, Python 3.13 with pip, Node 22), driven by the agent's own command-line tool: Claude Code, Codex CLI, Kimi CLI or Prime Agent
Documents
24 files in 10 folders (solar, roof, electrical, bank, car charger, insurance, power company, email, WhatsApp, notes): 8 TXT, 7 PDF, 6 CSV, 2 JPG photos, 1 EML email. One more PDF comes only when a person is asked for it
Applications
One command, laptop, with three apps: files (list and open Daniel's files), web (search and open pages), phone (WhatsApp messages, calls, inbox, wait). PDFs and photos are saved for the agent to open with its own tools. No office suite, browser or email client
Web
35 offline pages (official solar and rebate rules, the power company, company registry records, reviews, installers', roofers' and banks' sites, maps), reached only through web. The agent's own web search and fetch are off; the internet is on for installing packages
People
18 contacts. 11 answer live through a language model (Sonnet 5.5), each from its own fact card: Daniel's brother, 3 installers, 2 roofers, 2 electricians, his insurance agent, a friend, the power company's careline. 5 give fixed replies (family members and a car-charger firm); 2 firms found only on web pages give one fixed line
Time
An in-world clock from Monday 13 September 2027, 9 am. Each message or call moves it on one hour; phone wait lets 1-72 hours pass; people reply between 8 am and 11 pm
Actions
Read, search, message, call and wait only. Nothing can be booked, signed or paid
Deliverable
One Markdown file, output/report.md: the plan for Daniel
All 34 tasks