We gave an AI agent ten tasks in plain English — a code review with fixes, a Telegram bot with tests, a data report from a CSV, a few web apps — and ran all of them on one inexpensive model. We didn't write or edit a line of code. Ten tasks took 52 minutes of model time and $2.00 in total.
Task 1: fix someone else's code
A folder with a small Python to-do app written by someone else, plus its tests. Most tests failed. The prompt:
Act as a careful senior developer — find all the bugs, including the ones the tests don't catch, fix them and write a report.
The agent ran pytest, read the code, found 7 bugs, fixed them without touching the tests and wrote a report:
The agent's report: 6 failed / 1 passed before, all 7 passed after — $0.01
One of the bugs is the classic that tests rarely catch:
- def __init__(self, tasks=[]):
- self.tasks = tasks
+ def __init__(self, tasks=None):
+ self.tasks = list(tasks) if tasks is not None else []
Every TodoList() shared the same list — add a task to one, and it shows up in all of them. Another bug swapped day and month in a date format. Before: 6 failed, 1 passed. After: all 7 passed. One run, one minute, $0.01.
The setup
- one model for everything — Claude Haiku 5.5, one of the least expensive in its family;
- the agent can create files, run one program at a time (Python, Node, test runners), open a page in a clean browser window, take a screenshot and read the console errors;
- we only looked at results and, when needed, asked for changes;
- the price is the OpenRouter bill for all runs, fixes included.
The agent runs in Valaris, a desktop app (and web version) we build at BrainWave Studio: hundreds of models from OpenRouter in one chat, chats stored only on your machine.
All ten tasks
| Task | Time | Cost | Runs |
|---|---|---|---|
| Fixing someone else's code | 1 min | $0.01 | 1 |
| Telegram bot with a shop inside Telegram | 5 min | $0.10 | 2 |
| Sales report from a spreadsheet | 2 min | $0.07 | 1 |
| Crypto dashboard for Bybit | 4 min | $0.07 | 2 |
| Online store | 4 min | $0.07 | 1 |
| Coffee shop website in 5 languages | 7 min | $0.10 | 2 |
| Habit tracker app | 3 min | $0.05 | 1 |
| Solar System quiz for kids | 2 min | $0.03 | 2 |
| Picture book “Pip and the Lost Star” | 9 min | $0.19 | 4 |
| Browser game “Gale the Hedgehog” | 15 min | $1.31 | 5 |
| Total | 52 min | $2.00 |
"Time" is from sending the task to the finished result; "runs" is how many times we sent the task or asked for a change.
Task 2: a Telegram bot — with 34 tests nobody asked for
A flower shop as a Telegram Mini App: catalog, cart, promo code, checkout — $0.10
The bot plus a Mini App shop inside Telegram: catalog, cart, promo code, checkout, colors from the Telegram theme. Unprompted, the agent wrote 34 tests for the bot, and all of them pass. Five minutes, two runs, $0.10.
Task 3: a sales report from a CSV
848 orders → a Python script → a dashboard with charts and conclusions — $0.07
A CSV with 848 orders for a year. The agent wrote a Python script that computed revenue, average order value, top products and regions and the monthly trend, then turned the numbers into a dashboard with charts and plain-language conclusions. Two minutes, $0.07.
What we learned
- Running code beats reading code. The cheap model got the review right because it could run the tests, not because it "understood" everything on the first read.
- Screenshots catch what tests don't. For the web apps, opening the page in a clean browser window and looking at it fixed layout problems no test would flag.
- Cost follows back-and-forth, not difficulty. One-shot tasks cost cents. The browser game was the most expensive ($1.31) because we built it in five small steps.
- Sandboxing is the real feature. The agent only sees folders you picked; every file write asks first and shows a diff; any change can be undone; there is no delete; it runs one program at a time, never a shell, with a time limit and without your environment's keys and tokens.
The other seven
Crypto dashboard (read-only key via a local proxy, demo data) — $0.07
Online store: 16 products, filters, cart, 3-step checkout — $0.07
A coffee shop site in 5 languages — $0.10
Browser game — the most expensive, $1.31, built in 5 steps
What it doesn't replace
- These are small, self-contained projects. Real codebases have more connected parts and need more review.
- A human still decides when it's done.
- All companies, people and data in the projects are made up — it's a demo, not client work.
Try it
- All ten projects, live, with the exact prompts: valaris.app/portfolio
- The app (Windows) and the web version: valaris.app
- Every release with SHA-256 checksums and the changelog: github.com/ValarisAPP/valaris-releases
Valaris is free; agent mode is part of Valaris Pro, and models are paid at their OpenRouter price. What would you hand to a cheap model? Drop a task in the comments — we'll run the interesting ones and post the results.
— BrainWave Studio, the team behind Valaris
Top comments (0)