moshcode work: 50 agents, one verifier, and an inbox

Описание и применение

Мошкод реализует автоматизированный процесс проверки и исправления тестов с использованием 50 параллельных AI‑агентов, автоматически повторяя неудачные попытки и собирая успешные патчи.

Задача

Managing thousands of AI agents for math proofs requires a scalable verification workflow that isolates failures and aggregates results without overwhelming operators.

Результат

In the first run on moshcode, 139 local test failures were reduced to 0 by identifying 19 true failures and automatically fixing them with 50 parallel workers. The inbox remained empty, and no assertions were altered.

Рабочий процесс

• 1. Use moshcode work from-tests fix-tests to create a task per failing test file.
• 2. Run tasks with moshcode work run fix-tests -n 50 --engine claude to launch 50 parallel workers.
• 3. Check progress with moshcode work status fix-tests for aggregated numbers.
• 4. Inspect only failed attempts with moshcode work inbox fix-tests.
• 5. Apply passing patches using moshcode work apply fix-tests.
• 6. Repeat with additional verifiers such as typecheck or URL checks as needed.

Доказательства

Every task carries a verify command, and exit 0 is the only definition of done.

Источник

The inbox stayed empty. No assertion was added or removed.

Источник

A failed attempt goes back in the queue with the verify output in its next prompt.

Источник