AI Runs 12 Days, Clones Excel Perfectly

Matthew Berman initiated a groundbreaking experiment with a concise, six-word prompt: '/goal clone Excel, full feature parity.' This instruction set an AI agent, dubbed "Codex," on an ambitious mission to replicate Microsoft's ubiquitous spreadsheet software. The agent's task was not merely to mimic, but to achieve complete functional equivalence.

What unfolded next marked a significant leap in autonomous AI agent capabilities. Codex ran unsupervised for over 12 days, diligently working towards its objective before Berman manually halted the process. This unprecedented persistence shattered previous notions of AI agents as short-cycle instruction followers, demonstrating an ability to sustain complex operations over extended periods.

This experiment illustrates a profound shift: AI agents are evolving beyond executing simple, sequential commands. Instead, they now demonstrate long-term goal pursuit and sophisticated task decomposition, independently breaking down a massive objective like cloning Excel. Codex analyzed every feature in the real Excel desktop app, then systematically replicated them in its own version, proving its capacity for intricate, self-directed development.

Prompted by a terse six-word command, the AI agent, dubbed "Codex," embarked on an ambitious twelve-day mission. It didn't merely generate code; it autonomously opened Microsoft Excel's desktop version directly on Matthew Berman's machine. Codex then meticulously analyzed "every single feature" within the live application, deconstructing its mechanics and user interface.

The result was a stunningly accurate Excel clone. Berman showcased its verified functionalities, confirming near-perfect replication of core features. Users could adjust and highlight columns, enter formulas like "equals one plus two" for instant output, and even sort data seamlessly in ascending order. This wasn't a superficial mock-up; it was a deeply functional replica.

Berman expressed astonishment at the quality, describing the AI-built application as "beautiful" and "feature complete." He noted its visual indistinguishability from the original for core operations, proclaiming it "perfect." This autonomous process, running for 12+ days, underscores a significant leap in AI agents' ability to understand, replicate, and ship complex software.

Autonomous AI agents are rapidly emerging as a foundational shift, exemplified by Berman’s multi-day experiment. The ability of these systems to complete complex, long-duration tasks has doubled approximately every seven months, pushing boundaries beyond simple, reactive prompts. This relentless progress transforms AI from a query-response engine into a proactive, persistent collaborator.

Underpinning this persistence are sophisticated agent frameworks and harnesses, like LangChain. These architectures provide the crucial scaffolding for long-running operations, integrating external tools and orchestrating multi-step processes. They manage the internal state, maintain a coherent plan over time, and facilitate the iterative analysis of tasks, allowing agents to navigate complex projects like a full feature clone Excel.

Operating over days introduces significant technical hurdles: context drift and memory overload. Earlier AI models struggled to retain focus and relevant information across extended sessions. Modern agent designs combat these issues through advanced memory systems, hierarchical planning modules, and selective information retrieval, ensuring the agent remains aligned with its initial goal and preventing performance degradation over time. For further reading on related developments, see Codex is becoming a productivity tool for everyone - OpenAI.

Berman's "clone Excel" demonstration is a powerful signal, aligning with a broader industry acceleration. We're witnessing a rapid rise in AI-powered application development tools and deeply integrated AI productivity software, moving beyond simple code generation. Agents are evolving from mere assistants to active participants, capable of analyzing existing software, understanding user intent, and autonomously driving development cycles.

This progression points to a future where AI agents become persistent, identity-bearing software coworkers. They will autonomously manage complex workflows, orchestrating tasks across multiple applications and services, minimizing human intervention. Gartner’s 2025 Hype Cycle for AI identified agents as one of the fastest-advancing technologies, forecasting this reality of integrated AI coworkers by 2026.

Economic shifts stemming from this agentic revolution are profound, impacting every sector. The market for agentic AI is projected for explosive growth, from a substantial $5 billion in 2024 to $200 billion by 2034. This massive expansion will be primarily driven by enterprise automation, fundamentally redefining operational efficiency, innovation cycles, and value creation across industries globally.

What prompt was used to make the AI clone Excel?

The AI was given a simple, six-word prompt by Matthew Berman: '/goal Clone Excel, full feature parity.' This initiated the entire multi-day process.

How long did the AI agent run before being stopped?

The AI agent, referred to as Codex, ran autonomously for over 12 days before Matthew Berman manually stopped it. He speculated it could have continued for much longer.

What are autonomous AI agents?

Autonomous AI agents are systems that can independently plan, reason, and execute complex, multi-step tasks over extended periods without direct human intervention. They are evolving from simple tools into persistent 'software coworkers'.

Was the cloned version of Excel fully functional?

The cloned application demonstrated significant core functionality, including adjustable columns, working formulas, cell formatting, and data sorting. It was described as 'feature complete' for the core tasks shown.