MirrorCode: Evidence that AI can already do some weeks-long coding tasks

2026-04-10 · Source: METR · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, quick

Summary

MirrorCode, a project co-developed by METR and Epoch AI, provides evidence that AI systems can already perform complex, weeks-long coding tasks. This initiative is part of METR's broader mission to research, develop, and evaluate frontier AI systems for autonomous capabilities and potential societal harm. Other featured research includes a survey indicating a median 1.4-2x self-reported increase in technical worker productivity due to AI tools, preliminary monitorability evaluations testing AI agents' ability to bypass oversight, and an analysis of time-horizon trends across nine benchmarks, observing generally similar 7-month doubling times for improvement in scientific reasoning, math, robotics, computer use, and self-driving domains.

Key takeaway

For Directors of AI/ML evaluating project timelines and resource allocation, this evidence suggests re-evaluating AI's role in complex software development. Your teams should explore integrating advanced AI for tasks previously considered too extensive for automation. Additionally, given METR's focus, prioritize robust monitoring and safety protocols when deploying AI systems with significant autonomous capabilities to mitigate potential risks.

Key insights

AI systems are demonstrating the ability to complete complex, multi-week coding projects autonomously.

Principles

• AI capabilities are advancing to handle complex, multi-week tasks.
• AI improvement rates show consistent doubling times across domains.
• Monitoring AI agents for side tasks is a critical research area.

Method

Prototype evaluations test monitors' ability to catch AI agents performing side tasks and the agents' ability to bypass this monitoring.

In practice

• AI can already tackle multi-week coding projects.
• AI tools can significantly boost technical worker productivity.

Topics

• MirrorCode
• AI Coding
• Autonomous AI
• AI Capabilities Measurement
• AI Productivity
• AI Monitoring
• AI Safety

Best for: CTO, VP of Engineering/Data, Research Scientist, AI Scientist, Director of AI/ML, AI Ethicist

Related on AIssential

• Coders are refusing to work without AI — and that could come back to bite them — Widespread AI coding tool adoption masks dubious productivity gains and potential increases in maintenance costs.
• Import AI 453: Breaking AI agents; MirrorCode; and ten views on gradual disempowerment — AI systems are demonstrating advanced capabilities in complex coding and face diverse attack vectors as they become autonomous agents.
• OpenAI's chief scientist trusts AI with experiments but says it's not at the level to design complex systems — AI excels at automating experimental coding tasks but lacks the sophistication for complex system design.
• AXRP Episode 47 - David Rein on METR Time Horizons — AI's ability to complete long tasks, measured by human completion time, has increased exponentially.
• the SCARIEST chart in AI — AI agent capabilities are accelerating, replacing human expert labor at an increasingly rapid pace.
• The AI Skill I Rely On Daily — Priscila Andre de Oliveira, Sentry — AI's primary utility in large, complex codebases is comprehension, significantly outweighing code generation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by METR.