# Why We Switched From Claude Code to Codex

Podcast: AI & I
Published: May 7, 2026
Reading time: 18 min
Canonical: https://podbrew.app/briefs/ai-and-i-why-we-switched-from-claude-code-to-codex

Dan Shipper sits down with Austin Tedesco, Every's head of growth, to discuss a pivotal change in Austin's daily workflow. Austin now spends about 80 percent of his working time inside the Codex desktop app, using it for a wide range of tasks previously handled differently. This marks a significant transition from his earlier reliance on Claude Code.

The conversation explores why Codex has become Austin's daily driver. They examine how the agent management interface, a desktop app built on a coding agent, is transforming into a new operating system for knowledge work. This includes drafting go-to-market plans from meeting transcripts, rebuilding KPI dashboards, and managing communications.

This discussion underscores the transformative power of advanced AI agents in the realm of knowledge work. It illustrates how these tools are enhancing efficiency, streamlining complex tasks, and making sophisticated automation accessible, fundamentally reshaping how professionals approach their daily responsibilities.

## Key takeaways

- The rise of general-purpose coding agents that can access computer systems has created a new 'agent management interface,' functioning as a novel operating system for completing work.

- Despite finding Codex initially confusing and difficult, Austin continued using it for engineering tasks due to its effective results.

- The arrival of a new GPT model achieved performance parity with Claude Code's Opus, enabling a more balanced and specialized use of both AI tools in his workflow.

- The Codex desktop app serves as a primary daily driver, integrating with key platforms like Gmail, Slack, Notion, and Stripe to automate tasks and manage data efficiently.

- Despite the technical ease of switching between AI agents (as agents can help with data transfer), users often face an "emotional hurdle" when considering a transition due to perceived relearning effort.

- Custom reviewer agents within Codex can be tailored for strategic alignment and data accuracy, offering targeted feedback on plans rather than generic engineering reviews.

- AI models like Codex can effectively brainstorm practical automations by analyzing a user's specific workflow and frequently used tools such as Notion, Slack, and Gmail.

- Automations created by these models are often highly functional, requiring minimal adjustments to be immediately useful for daily tasks.

- AI agents can significantly streamline knowledge work by automating tasks such as compiling unresponded items and drafting responses, needing only simple user approval.

- Implement a system for auditing AI automations and prompting models to report their actions, ensuring continued effectiveness and preventing errors.

- General AI agents can streamline the creation and deployment of specialized agents through conversational interfaces, eliminating the need for manual configuration and traditional onboarding.

- Implementing a multi-agent strategy, where several specialized agents handle specific functions, is often more effective than relying on a single, all-encompassing master agent.

- AI models can synthesize complex go-to-market plans from disparate company data (meeting transcripts, Slack, templates), producing 80-90% complete drafts.

- Designing documents for both human readability and AI agent processing ('agent-first documents') enhances efficiency, allowing other team members and their agents to quickly understand and act on the plans.

- Implementing explicit rules and workflows, alongside a mandatory final human review, is critical for ensuring AI-generated content aligns with context and maintains quality.

- AI tools can efficiently transform existing human thoughts, such as those from meetings or dictated notes, into structured documents, allowing individuals to save time on formatting and concentrate on the intellectual work itself.

- Absolute data accuracy remains a significant challenge for AI-generated dashboards; critical metrics like MRR cannot tolerate even small percentage errors, necessitating meticulous human verification.

- Even with the requirement for manual data verification, AI-powered solutions allow non-engineers to independently build and manage complex data infrastructure, reducing reliance on specialized consultants or engineering teams.

- Workflows include a "compound step" where learnings from any session can be saved to a team-wide shared knowledge base and new workflows can be converted into automated skills for future use.

- Cultivating a company culture that values 'playing around' with new tools is essential because rapidly evolving workflows require continuous learning to avoid being outpaced by those adopting new paradigms.

## 00:00 - 06:02 Codex Transforms into an Agent-First Operating System for Knowledge Work

Codex initially served as a difficult, specialized tool for senior engineers focused on pair programming. OpenAI's early strategy centered on hobbling the model, anticipating general 'vibe coding' in ChatGPT. However, Anthropic demonstrated the value of a fast, smart, and emotionally intelligent model directly on a computer with file system access, providing a superior experience for programmers.

This shift in perspective revealed that if an agent can write software, it can effectively perform any knowledge work. Codex subsequently pivoted from an engineering-only tool to a daily driver for diverse tasks, including deep engineering, writing, and recruiting. It now integrates with services like Gmail, Slack, Notion, and Stripe, functioning as a comprehensive assistant.

This evolution signifies the emergence of a new 'agent management interface' acting as an operating system for how and where work gets done. This development has intensified competition among major AI companies. OpenAI offers Codex, Anthropic provides Claude Code and CoWork, and XAI acquired Cursor, all racing to deliver desktop agent apps that merge programming capabilities with general knowledge work functions.

In this agent-first world, the agent becomes the primary interface for interacting with software and the internet. This paradigm shift enables users to delegate complex tasks, send agents to interact with other software, and unlock new possibilities for productivity that were previously unattainable.

> there's a new operating system for how and where you're gonna get your work done, and it's this kind of agent management interface

## 06:02 - 10:02 Austin Tedesco's Conversion to Codex for Enhanced Workflow Automation

Austin Tedesco had his "agent pill" moment in December/January, diving deep into Claude Code through the command-line interface. He integrated it extensively into his work and personal life, finding it highly effective for automating tasks and acting as a thought partner. This made it essential for knowledge work that required strategic thinking, data analysis, and crafting marketing copy.

Despite initial nudges, Austin's first experience trying to build a "vibe-coded app" in Codex was challenging. He admitted that Codex made him feel "stupid" because of its complex questions and the need for constant clarification. While he persisted and appreciated the results for engineering tasks, about 80% of his general usage remained with Claude Code in the CLI.

A month ago, the introduction of the new GPT model brought a significant change. Austin observed a performance parity between the latest Opus model (Claude Code) and the latest GPT model (Codex) for his type of knowledge work. This balance allows him to leverage the specific strengths of each tool, influencing his overall workflow.

This shift highlights the evolving capabilities of AI agents, moving beyond single-tool reliance to a more integrated approach where users can select the best tool for specific tasks, even after an initial steep learning curve.

> nothing has ever made me feel more stupid than Codex like two months ago.

## 10:02 - 14:05 Codex as an Integrated Daily Driver and the Migration Experience

Austin Tedesco extensively uses the Codex desktop app, considering it his daily driver due to its speed and power compared to cloud-based alternatives like the Claude app. He describes it as the first tool he opens each day, integrating seamlessly with platforms such as Gmail, Slack, Notion, and Stripe to manage all his data sources.

Codex automates complex tasks; for example, Austin simply messaged it to "make the run of show" for a camp, and it instantly pulled relevant information from prior conversations, pushed the document to Notion, and sent it to Slack. He estimates spending 80% of his work time within the app because of its robust performance and the model's capabilities.

The AI agent market is currently a "horse race," with companies like Anthropic and others continuously releasing updates that push competitors towards parity or better. Despite the rapid advancements, switching between these tools is relatively easy, as agents can assist in transferring data.

However, an "emotional hurdle" exists for knowledge workers who are already comfortable with tools like the Claude desktop app, even if a new alternative offers significant improvements. The resistance stems from the perceived effort of learning a new system, despite the technical migration often being straightforward.

> When I tell them that I have fully transitioned to Codex, this like look of horror shows up on their face, and they're like, 'Do I...?' They're kind of like, 'Do I really have to?'

## 14:05 - 17:05 Configuring Codex for engineering and strategic workflows

Austin highlights Codex's efficiency for engineering tasks, such as shipping pull requests, noting its superiority over other tools like the Cloud Desktop app which he found clunky and slow during stress tests for go-to-market plans and PRs. He emphasizes that Codex handles these complex operations quickly and effectively.

He demonstrates his "EveryGrowth OS" setup within Codex, which is a folder containing essential secrets, API keys, and project instructional files that define the business and workflow. This centralized hub ensures all necessary resources are connected and easily accessible for the "Every" business.

A key feature of his setup includes custom reviewer agents within the EveryGrowth OS folder. These agents are designed to align with compound engineering principles but are specifically forked to focus on strategic alignment with company goals and data accuracy. This allows for targeted, model-driven reviews when developing plans, which is more relevant than general security reviews for strategic initiatives. He also mentions preparing to share his recommended prompts for new users to get started with Codex effectively.

> when you do stuff like that inside of Codex, it just works. Like, it just works really quickly and, and well

## 18:04 - 22:06 Brainstorming and Building Effective AI Automations with Codex

Austin describes his process for using Codex to brainstorm and implement automations within his daily workflow. He initiates a 'compound engineering brainstorm workflow' by prompting Codex to analyze his most frequently used tools, specifically Notion, Slack, and Gmail, and then suggest relevant automations to enhance his work.

Codex responds with intelligent suggestions like a 'follow-up radar' to manage information across various sources and a 'command center' for organizing events and camps, handling tasks such as pipeline management and tracking. This approach leverages the model's ability to identify pain points and propose practical solutions without the user having to pre-conceive the automations.

A key takeaway is the impressive functionality of the automations built by Codex. Austin notes that these automations work 'incredibly well' and require 'very little tweaking' to be integrated into daily use. He finds that they often 'just work' from the outset, providing immediate value.

As an example, Austin mentions an end-of-day automation that compiles all unresponded items, drafts replies, and can be actioned with a simple Slack reaction, effectively streamlining communication. He emphasizes that starting with a brainstorming automation state is an excellent way for knowledge workers to discover what AI agents can do for them.

> I do find that they just work incredibly well. They require very little tweaking to be like, 'This is a thing I would and do use every day.'

## 22:06 - 28:06 Implementing Human Review and Auditing for AI-Generated Outputs

When using AI to draft communications like Slack messages or emails, it's beneficial to conduct the final review in the native application (e.g., Slack or Gmail) rather than within the AI drafting tool. This approach provides a mental separation from the agentic space, allowing for a fresh perspective to ensure the content is exactly what you intend to send to a human. This method also applies to strategic planning documents, ensuring a final human check before engagement.

For advanced email triage, instead of directly instructing an AI agent on rules, it's more effective to have the agent "interview" you. This involves using a brain dump via speech-to-text to describe your current email problems and desired outcomes. The AI agent can then propose a comprehensive plan for filtering, summarizing, or archiving emails, which you can review and adjust to prevent the loss of critical communications, such as those related to potential earnings.

Crucially, regularly audit new AI automations to confirm they are working as expected and not creating unintended issues. Set reminders to check the automation's performance after a specific period, for example, 72 hours. Proactively prompt the AI model to report what actions it has taken, such as what emails it has archived, to verify accuracy and prevent missed opportunities.

> I always find that I get a better result rather than saying what I think the rules should be, and so I'll, I'll do a brain dump using Monologue or Speech to Text. App saying, 'Here's the problem I'm facing, my email's a mess, let's figure out how to triage it.'

## 28:06 - 32:08 Simplifying Specialized AI Agent Deployment with General Agents

General-purpose AI agents like Codex are changing how specialized agents are set up and deployed. Instead of manual configuration or dashboard settings, users can conversationally onboard specialized agents. The general agent leverages existing context from a user's computer and past interactions, automatically configuring new agents without traditional onboarding experiences.

This approach is inspired by strategies like Claire Vo's, who found significant breakthroughs by utilizing a suite of six specialized agents rather than trying to make one 'master' agent handle everything. The idea is that specific tasks are better served by agents tailored for those functions, leading to more effective and manageable AI systems.

For example, Austin Tedesco used Codex to design a suite of six specialized agents for his company's Slack. He provided Codex with the transcript of Claire Vo's interview and asked it to plan the agents, considering his work and the potential for Notion custom agents. This process moved from manual ideation to AI-assisted planning and provisioning.

While these specialized agents occasionally break, the general agent can also be used to diagnose and fix issues. Users can point Codex to problematic conversations, and it can intelligently adjust the architecture of the agents to resolve the problems, demonstrating a dynamic and responsive agent ecosystem.

> I actually just sent it the transcript of Claire's interview with Lenny and said, 'Like, I wanna do this too. Given everything you know about me and my work, make a plan to suggest six agents that we should provision into our Slack.'

## 32:08 - 36:08 AI Streamlines Go-to-Market Plan Creation for Humans and Agents

Austin Tedesco uses an AI called Codex to generate comprehensive go-to-market plans by feeding it existing company data. This includes meeting transcripts stored in Notion, discussions from Slack, and pre-defined plan templates. The AI synthesizes this information to create an initial draft of the plan, which Austin estimates is 80 to 90 percent complete.

The process involves an iterative approach where Austin nudges the AI and provides additional context. For instance, he reminds Codex to incorporate information about upcoming posts and launches from the company calendar, which it often overlooks initially. This refinement allows the AI to produce a proof document that consolidates all relevant strategic thinking.

A key aspect of this workflow is designing documents for both humans and AI agents. Austin explains that he doesn't just create these plans for human readability; they are also structured for agents to understand. This allows other team members, like Brandon, the COO, to use their own AI agents to quickly summarize the plan, understand the business case, or even assist with pricing modeling based on the generated document.

This AI-driven approach significantly improves efficiency and work quality. What previously would have taken Austin a full day or late-night hours to draft manually, is now handled rapidly by the AI. This transformation makes the strategic planning process faster, produces better outcomes, and creates a more engaging workflow.

> I don't make this plan for humans, I make this plan for humans and agents, and primarily for humans to understand through agents.

## 36:08 - 40:09 Ensuring Human Accountability and Thought in AI-Generated Content

While AI-assisted writing can be more efficient, its value hinges on the human author's willingness to stand behind the content. Many people would prefer to read AI-generated text if they know the author has thoroughly considered the ideas and can discuss them clearly. This contrasts with situations where AI is used without proper oversight, leading to a decrease in quality because individuals don't fully endorse what the AI produces.

To maintain quality and trust, it's essential to establish clear guardrails and workflows for AI usage. This includes setting specific rules within projects, such as prohibiting the AI from adding new information not previously stated or directing suggestions to a chat rather than directly into a document. A crucial step is always conducting a final human review before content is shared, ensuring everything aligns with expectations and accuracy.

AI can be particularly valuable for organizing existing thoughts into a readable format. Much of work involves translating already formed ideas into structured documents. For tasks like crafting company strategy or retrospectives, AI can convert spoken monologues or scattered notes into cohesive documents. This saves significant time on formatting, allowing individuals to focus on the core thinking and enabling them to tackle substantial tasks even in small pockets of time throughout the day.

> A lot of the time that you spend working is about taking thinking you've already done and putting it into a form that other people can read and consume, and the important part is doing the thinking.

## 40:09 - 44:09 Human oversight ensures accuracy in AI-generated KPI dashboards

Austin Tedesco is rebuilding Every's weekly KPI tracker in Notion, aiming to create a single, up-to-date dashboard accessible to both humans and AI agents. This dashboard tracks various business metrics like paid subscription trials, page views, and Monologue iOS MRR, integrating data from sources like Stripe using Notion's Workers tool.

The primary challenge in building this AI-powered dashboard is ensuring absolute data accuracy. Initial attempts with AI tools often produced numbers 5-10% off, which is unacceptable for critical metrics like Monthly Recurring Revenue (MRR). Running a business requires a source of truth that is exactly right, not even 3% off.

To address this, Austin manually verifies each column of the AI-generated data, ensuring it is "exactly right and defensible." This meticulous human oversight is essential to reliably run and grow the business and confidently deploy AI agents to act on the data.

Despite the need for manual verification, using AI tools still significantly streamlines the process. It enables a non-engineer to build and maintain a complex, always-on dashboard that updates every six hours, bypassing the need to hire consultants or pull from engineering resources.

> Our MRR number can't be five percent off. Like we can't run a business with a source of truth that's even three percent off. It has to be just exactly right.

## 44:09 - 46:30 AI Streamlines Recruiting by Identifying Niche Candidates, While MRR Measurement Remains Philosophical

Determining how much money a company has made and grown, specifically its Monthly Recurring Revenue (MRR), is not an objective calculation but a philosophical question. Businesses must decide on a consistent method for measurement, as there is no single universally correct way to define these figures.

Beyond financial philosophy, AI presents practical applications in knowledge work, such as recruiting. When seeking a Head of L&D, a company wanted to find individuals with experience running high-quality technology courses, particularly those from General Assembly during its peak in the 2010s.

Leveraging an AI tool like Codex, the team requested a list of General Assembly alumni who had subsequently transitioned into the field of AI. The AI quickly provided relevant candidates.

The first candidate identified by Codex was a perfect fit, leading to a direct outreach via Twitter. This demonstrated AI's power to efficiently pinpoint highly specific candidate profiles, a significant improvement over sifting through numerous general applications.

> it was just one of those holy shit light bulb moments where normally what we're doing is sorting through a ton of applications

## 50:09 - 54:10 Austin Tedesco explains how he forked the Compound Engineering plugin to create "Compound Knowledge" for general knowledge work and how to save session learnings as automated skills.

Austin Tedesco adapted the Compound Engineering plugin, which was designed for specific engineering tasks like reviewing security and frontend design, to create "Compound Knowledge." This new version is tailored for general knowledge work, focusing on aspects like strategic alignment and data accuracy. He found it to be a valuable learning experience in developing custom plugins.

The Compound Knowledge plugin allows users to customize review processes. For instance, a social media marketer could configure it to ensure all reviews adhere to their specific style guide and past performance metrics. The original Compound Engineering plugin also works effectively for general knowledge work right out of the box, even without modifications.

A key part of Austin's workflow involves compounding knowledge: after any session in tools like Codex or Cloud Code, agents are instructed to ask if the learnings should be saved to a team-wide shared compound source of truth. This ensures that valuable insights are captured and made accessible across the team.

Additionally, the system prompts users to consider turning any new workflow discovered during a session into an automated skill. This allows for repeated tasks to be performed automatically, streamlining future work and increasing efficiency.

## 54:10 - 58:11 Fostering AI Adoption Through Culture and Dedicated Experimentation

Balancing the pursuit of innovative AI experimentation with the demands of essential daily business tasks presents a significant challenge for many organizations. While powerful AI models are exciting and can lead individuals to over-index on learning, the core work still needs to get done. Every addresses this by integrating a culture of 'playing around' directly into its operational philosophy, a practice actively encouraged by leadership.

This culture recognizes that the landscape of tools and workflows is evolving at an unprecedented pace. Simply focusing on current job functions and existing processes risks being outmaneuvered by others adopting new tools and paradigms. Dedicating time to experiment, even if it initially feels like a diversion, is seen as crucial for leveling up skills and staying competitive in a rapidly changing environment.

To formalize this approach, Every implements specific organizational practices. One notable example is 'Think Week,' held twice a year, during which employees completely pause day-to-day work. This dedicated week is used for playing with new technologies, building experimental tools, and learning collectively. This focused time provides the necessary space for deep exploration that casual daily experimentation might not allow.

> if you just give yourself some time to play around, it may feel like a waste of time, but you're leveling yourself up to a different game at a different level, and I think that's really important.

---

Get podcast briefs for shows you follow: https://podbrew.app/