OpenAI’s Plan to Make ChatGPT the Everything App — Akshay Nathan, OpenAI
Akshay Nathan, head of Core Product Engineering at OpenAI, discusses the journey and rationale behind ChatGPT Work. He explains how Codex's unexpected adoption by non-developers led to a unified agent harness, blurring the lines between developer and knowledge worker tools. The conversation delves into model capabilities, the role of artifacts and interactive 'Sites' in replacing traditional documents, and how AI fosters a new era of 'T-shaped' generalists. Nathan emphasizes the shift in productivity bottlenecks to 'ideas and taste' and the importance of 'quality at-bats' over mere 'motion' in an AI-powered development landscape.
Akshay Nathan, leading Core Product Engineering at OpenAI, shares insights into the evolution of AI products, particularly ChatGPT Work, emphasizing his long-standing vision of making software power accessible without code. His journey from no-code platforms like Airtable led him to recognize LLMs as the missing piece to achieve this.
The Genesis of ChatGPT Work
Akshay explains that ChatGPT Work emerged from unexpected internal observations at OpenAI. The Codex agent, initially designed for developers, saw a significant uptake among non-developers (e.g., strategic finance, marketing). Users reported feeling they had a “superpower,” demonstrating that the power of agents extended beyond traditional coding roles. This realization spurred OpenAI to merge these agentic experiences into ChatGPT Work, aiming to bring this capability to their large ChatGPT user base.
Codex vs. ChatGPT Work: A Shared Foundation
While Codex and ChatGPT Work present different user experiences, they share the same underlying agent harness. Improvements to this harness, particularly for knowledge work, include enhanced support for plugins, computer use, and artifacts. The distinctions primarily lie in UX: Codex exposes more Git-related visibility and a “diff-forward” experience, while ChatGPT Work abstracts these details. Safety considerations, such as sandboxing defaults, also differ. The overarching goal is to avoid forcing users to choose an experience, ensuring that capabilities are largely interchangeable.
Merging Agent Experiences and Blurring Job Functions
OpenAI made a deliberate decision to merge agent experiences rather than maintain separate products. Akshay posits that AI is rapidly blurring the lines between traditional job functions—engineering, design, strategy, operations. By removing rigid boundaries in their tools, OpenAI empowers users to leverage AI regardless of their role. This is reflected in unified primitives like plugins, supporting a future where individuals are less siloed by their job titles.
Model Capabilities and User Control
The continuous advancement of models is critical for these products. OpenAI strives for an opinionated default model configuration that is “the best possible” for most users. However, for power users, options exist to adjust reasoning levels, utilize “Ultra” mode, or engage multi-agent setups, particularly for complex or highly parallelizable tasks. The UI includes a slider to intuitively balance speed/efficiency against quality/thoroughness.
Artifacts, Sites, and New Forms of Work Product
A significant focus of the launch was improving artifacts and introducing “Sites.” Akshay highlights that Sites could potentially replace traditional documents like decks and spreadsheets for collaborative knowledge work. Because Sites are essentially HTML, they offer infinite flexibility and higher bandwidth for information sharing. Internally, teams are already adopting Sites as their primary mechanism for ongoing reports, enabling richer, more interactive communication.
Designing for a Generalist Agent
The core challenge in designing ChatGPT Work is balancing simplicity with its immense capability. The product aims to provide sufficient UI for users to express their needs, verify tool usage, and source data, but then “gets out of the way.” A key strategy is “show not tell,” enabling users to discover new use cases through experience rather than explicit instruction.
The Path from Developer Agents to Everyone
OpenAI envisions a phased rollout: starting with developers (early adopters), then expanding to general knowledge workers, and finally to everyone in their daily lives. Akshay shares personal examples of using ChatGPT Work for meal planning, workout tracking, and even package retrieval, framing ChatGPT as a potential “personal operating system” for life management. He notes the fluid boundary where personal use cases often inform professional ones.
Power User Advice and AI-Assisted Performance Reviews
Akshay’s advice for power users includes: (1) Broaden your imagination: Models improve rapidly, so re-evaluate tasks previously thought impossible. (2) Provide more context: Connecting plugins, local files, and leveraging persistent memory allows the agent to become more valuable and proactively insightful. He illustrates this with performance reviews: AI is not for generating the review itself but for “agentic search,” gathering comprehensive context from code, Slack, and documents to aid human judgment, a task that was not effectively possible just six months prior. He also recounts an internal OpenAI automation that scans Slack and documents to generate memes, showcasing the models’ surprising ability to find connections and even humor.
OpenClaw’s Influence and ChatGPT Memory
OpenClaw, an independent personal agent project, provided significant inspiration for features within ChatGPT Work, such as persistent computer environments and scheduled tasks. OpenAI aims to bring the “magic” of such personal agents to a wider audience through ChatGPT. Furthermore, ChatGPT’s “Memory V3” system allows the model to learn about users over time, creating a personalized experience where users feel their ChatGPT “is their ChatGPT.” Chronicle, an experimental feature, further enhances memory by learning from computer usage, building deeper context for proactive assistance.
AI’s Impact on Product Development and Productivity
AI is profoundly transforming product development, accelerating the journey from idea to execution. Traditional roles are blurring, leading to a future where “everyone will be T-shaped”—generalists empowered by AI, retaining deep specialties. This shift makes “ideas and taste” the new bottlenecks, as the ability to build becomes widespread. Akshay notes LLMs still struggle with genuinely “new ideas,” which typically stem from user feedback and real-world friction. Measuring productivity is also evolving; old proxies like code commits or story points are becoming less relevant. Instead, the focus shifts to “quality at-bats”: a team’s efficiency in iterating from idea, to build, to feedback, and validation. A critical trap to avoid is “conflating motion and progress”—AI can generate much motion, but true progress requires deliberate and prescriptive goal-setting.