Codex for Everyday Work: AI Agents Beyond Coding
Thibault Sottiaux, Head of Codex at OpenAI, discusses the evolution of Codex from a niche developer tool into a general-purpose agent for knowledge work. He explores how this shift is redefining productivity, team dynamics, and our fundamental relationship with technology.
Thibault Sottiaux, the head of Codex at OpenAI, explains that the journey of Codex began with the grand challenge of creating a model that could code at the level of a prolific software engineer. The first public version, “Codex web,” was a cloud-based interface that would read a task, access a code repository, and open a pull request on GitHub with the proposed changes. However, this approach had significant friction; it was difficult to set up, and the models weren’t yet reliable enough for long-horizon tasks, making iteration difficult. This led to a pivot towards a more integrated, local approach.
The Shift to a General-Purpose Agent
The realization that Codex could be more than a coding tool emerged recently, catalyzed by a step-change in model generality and reliability. The team observed that even software engineers spend only about 20-30% of their time writing code. The majority of their work involves non-coding tasks: prioritizing tickets, architectural discussions, investigating bug reports, managing on-call duties, and gathering information.
Technical users naturally started using Codex for these adjacent tasks. The agent became significantly more useful when it was given access to more context beyond code, such as documents and information in Notion. Today, the majority of tasks performed in Codex are non-coding tasks.
A key “aha” moment came during a product launch when a product manager, Alexander Amirus, used Codex to manage the entire process. He delegated the work of tracking changes, chasing up team members via Slack, and coalescing user feedback into an up-to-date plan—all while he was in a meeting. This demonstrated a profound shift in productivity, where the agent handled the time-consuming coordination work that typically bogs down teams.
Redefining Workflows and Roles
The adoption of Codex at OpenAI has had a ripple effect across the organization, accelerating not just engineering but also adjacent roles like design and product management. The bottlenecks in productivity have shifted to other departments, such as communications and marketing.
This acceleration is changing the nature of work itself:
- Empowerment: Individuals are empowered to solve their own problems. For example, anyone can now query business data directly without waiting in a queue for the data analyst team. This closes the loop between the person with the problem and the person who can build the solution.
- Blurring Roles: Designers are now pushing code changes themselves, allowing them to shape the product experience directly without needing to convince an engineering team to prioritize minor-seeming but impactful changes.
- Personalized Software: We are entering an era of “home-cooked personalized software.” Thibault demonstrated this by creating a custom webpage with a map of the best bread loaves in San Francisco, complete with prices and descriptions, generated from a single voice prompt in minutes. This is a task that would have previously taken a skilled person a weekend, if it were attempted at all.
A New Relationship with Technology
Codex is evolving into a “trustworthy partner” that can manage tedious work and reduce cognitive load. Thibault uses it as a personal chief of staff, delegating over a hundred tasks a day, from organizing his desktop files to generating a daily summary of priorities from his email, calendar, and documents.
This addresses a core problem of modern knowledge work: information overwhelm and burnout. Instead of being trapped by our tools, we can be liberated by them. The agent acts as a shield, filtering noise and flagging only what is important.
The future vision is to move beyond turn-based, task-oriented interactions to a continuously running agent that works proactively, 24/7, to help you achieve your goals.
The Future: Long-Horizon Tasks and /goal
A new advanced feature, /goal, allows users to give Codex a long-term objective that it can work on relentlessly for hours, days, or even weeks. This is being used for complex problems like solving hard mathematical proofs, rewriting entire programs, and achieving scientific breakthroughs. The key to success with /goal is to be precise in defining what success looks like, much like giving clear instructions to a new team member.
Enterprise Adoption and Trust
For broader enterprise adoption, the primary bottleneck is not model capability but trust. Organizations need assurance that the agent is safe and secure. OpenAI is addressing this through several mechanisms:
- Sandboxing: By default, agents run in sandboxed environments with tight controls, restricting file system access to specific folders and disabling network access if needed.
- Enterprise Controls: Fine-grained permissions allow administrators to control what the agent can and cannot do (e.g., granting read-only access to sensitive data).
- Auto Review: A novel safety feature where a second “referee” agent reviews the actions of the primary agent in real-time. It is incentivized to flag and stop any action it deems too risky, providing a critical layer of oversight.
The most common mistake users make is delegating their own understanding along with the task. It’s crucial to use the agent not just to do work, but also to learn and get explanations, ensuring the user remains grounded and in control.