How Codex Became a Directionless Mess

Summary

Ben Davis argues that Codex began as a strong desktop coding agent but became less coherent after it was folded into the ChatGPT desktop app. Although individual capabilities have improved, he says the overall experience now feels like several products joined together without a consistent design.

Main Product Criticisms

• Performance and interface: Chat threads can load slowly, especially in older profiles, and the interface has numerous small usability problems.
• Priority sidebar: Davis says Codex’s “priority” view behaves more like a timeline of recent work than a useful queue of issues requiring attention. He contrasts it with T3 Code’s sidebar, which he considers better suited to parallel work: review an issue, address it, dismiss it, and move on.
• Codex and ChatGPT are split awkwardly: He understands the idea of offering a simpler ChatGPT experience alongside Codex, but thinks the boundary is confusing. His preference is one underlying agent that can answer everyday questions or handle coding tasks as needed.
• “Dots” and persistent work: Davis finds the shared-bot approach limiting for organizing long-term work across separate areas such as business, finances, projects, and personal tasks. He favors distinct, named agents with their own context and routines.
• Spaces: He uses Spaces to organize repeatable YouTube workflows, including editing instructions, thumbnail prompts, upload checklists, and templates. He describes it as something like Notion with AI features, but says its connections to Codex threads are confusing. For example, making changes can pull the interface back to a particular Space page, and threads appear both as regular Codex branches and as items within Spaces.

Broader Analysis and Comparisons

Davis’s central explanation is that strong models can excel at tasks with clear ways to verify success, such as making tests pass. Product design, however, requires subjective judgment, prioritization, and taste. Those decisions are difficult to capture in a validation function, so rapid AI-assisted development can produce many features without a cohesive product vision. He calls the resulting accumulation of disconnected features a form of product “decay.”

He compares OpenAI’s products with several alternatives:

• Claude Code: He praises its CLI, while noting that it took time to mature.
• T3 Code: He says its sidebar is better for his workflow and presents it as an example of a product improving through clear, consistent product decisions.
• Grok Bot: He describes it as polished and consistent across desktop, mobile, web, and other services. He says its underlying structure resembles Cursor cloud agents, with separate streams and cloud environments. In his view, he chose it over a competing option with a stronger model because the product experience was better.
• OpenAI’s announcements: He says Dev Day introduced so many products and features that useful releases—including the Codex CLI—were easily lost among the announcements. He considers the CLI strong and says he uses it frequently. He also calls Spaces useful, but criticizes Codex’s cloud offering as behind alternatives such as Cursor and T3 Code.

Suggested Fixes

Davis would simplify Codex before adding more features:

1. Make a Codex stream the common foundation for the CLI, desktop, and web experiences.
2. Treat “Dots” as named streams with attached memory, routines, and connections.
3. Make it easy to run a stream locally or in the cloud.
4. Refine Spaces into a more independent building block, rather than tightly binding their composer to a particular page or thread.
5. Focus on a clear product direction and remove features that do not support it.

He believes OpenAI has the talent to improve the product, but says the underlying structure needs to be fixed first.

Sponsor Mention

The video includes a Firecrawl promotion. Davis describes its web-data tools as available through MCP, CLI, API, and SDK, and highlights Alexandria, a set of curated data sources. He demonstrates using it to gather information about recent San Francisco startup activity.

Reviews, Guides, and Demonstrations

• Hands-on review: Critique of the ChatGPT desktop and Codex experience, including performance, navigation, and product organization.
• Feature walkthrough: A demonstration of using Spaces to organize repeatable video-production workflows.
• Recommendations: Proposed changes to Codex’s architecture and the relationship between streams, agents, and Spaces.
• Tutorial: No separate step-by-step tutorial is presented.

Main Speakers and Sources

• Main speaker: Ben Davis, providing the product review and analysis.
• Products and companies discussed: OpenAI Codex and ChatGPT, Anthropic and Claude Code, T3 Code, Grok Bot, and Cursor.
• Sponsor: Firecrawl.