# In a New Benchmark, an AI Agent Rebuilds a 60,000-Line Program From Its Behavior Alone

Epoch AI and METR's MirrorCode measures whether models can reconstruct working software they cannot read, and the early scores are higher than program-synthesis researchers expected.

- Published: 2026-06-28T10:47:24.054Z
- Canonical: https://polylog.news/ai/2026-06-28/in-a-new-benchmark-an-ai-agent-rebuilds-a-60-000-line-progra
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [Epoch AI](https://epoch.ai/blog/mirrorcode-preliminary-results), [METR](https://metr.org/blog/2026-04-10-mirrorcode-preliminary-results/)

Epoch AI, working with METR, has released [MirrorCode](https://epoch.ai/blog/mirrorcode-preliminary-results), a benchmark that tests whether a model can rebuild a complete application without seeing its source code. The agent receives a compiled binary it cannot read, along with documentation and a test suite, then must reproduce the program so that its output matches the original exactly, including on held-out tests. The framing, as the Russian-language channel AI ML Big Data put it, is direct: [can large language models rewrite software from scratch?](https://t.me/ai_machinelearning_big_data/10419)

The main result is that Claude Opus 4.7 scored 56 percent across 25 target programs spanning Unix utilities, serialization tools, bioinformatics, interpreters, and cryptography. On one task, reimplementing a 16,000-line bioinformatics toolkit, it passed 99.95 percent of tests, work the team estimates would take a human engineer between two and seventeen weeks. The model finished in roughly 14 hours for about $251. It also reconstructed pkl, a configuration language of roughly 60,000 lines, [the largest autonomous coding result documented in a public evaluation](https://metr.org/blog/2026-04-10-mirrorcode-preliminary-results/).

The balanced reading lies in what the benchmark does and does not show. It confirms that agents can now sustain goal-directed work across task horizons measured in days, not minutes, on small-to-medium codebases with clear automated tests that define correct output (test oracles). It also confirms the ceiling. On large projects, success drops sharply, and MirrorCode's design favors programs with deterministic, fully specified behavior, the easiest case for copying a program's behavior. Real software, with ambiguous requirements and no exhaustive test suite, remains beyond current models.

## What this means

A 56 percent score on week-scale coding tasks is evidence that the frontier of autonomous coding has moved from snippets to whole programs. The result is bounded by exactly the conditions that rarely hold in production: a complete test oracle and a clean specification. For engineering organizations, the near-term value is in reconstruction and migration tasks, where behavior is observable and testable, not in building new systems from scratch. The benchmark also gives the oversight debate a concrete number to track as models improve on it.

## What to watch

- How successor models score on MirrorCode's large-project tier, since the binding limit is project size, not whether the approach works at all.
- Whether the exact-match-on-tests methodology is examined for behavioral-cloning shortcuts, which would show whether the scores reflect genuine reconstruction or test-suite overfitting.