Own your ~~data~~ workflow
Making sense of coding agents: Claude Code and now OpenAI Codex (strategy from an AI company perspective)
Why did these companies build, released and now are hyping these tools? They want you to be productive, use these tools to build better services and beat them in software development game. Yeah, that must be it.
Internet was fully scrapped and now these companies onto building themselves input streams of "quality workflow" user data.
Both Anthropic and OpenAI obviously collect convos from the LLM chat interfaces they offer and leverage that data. Think about what kind of data you as a developer / analyst / <name yourself> provide now via these new tools. You ask the tool to tackle a problem, feed your codebase to it, as it does you monitor the progress, as LLM thinks you catch and identify errors or hallucinations it made and then you instruct to correct it.
There are 3 components here: existing project files (or a project from scratch), LLMs thought process, and your (user's) continued guidance via natural language where if something goes wrong you tell (teach) LLM a correct way.
Thanks for free guidance. And ..., a developer also pays for the tokens used. We pay to give quality use cases to these models. Sweet.
Something's fun
Brief history: Anthropic released Claude Code as a proprietary closed-source tool. 2 months later OpenAI releases Codex, an exact copycat of Claude Code in terms of UI, open-source.
Why Claude Code is closed source? The drama about it is real: 1) https://github.com/Aider-AI/aider/issues/3362#issuecomment-2681023123 2) multiple developers took it for a swim, published their research on gh: https://github.com/mandeep511/OpenAGI/, https://github.com/dnakov/claude-code 3) Anthropic demanded Github to take them down: https://github.com/github/dmca/blob/master/2025/03/2025-03-10-anthropic-2.md 4) Anthropic packed shitload of prompts into Claude Code - is this why it is closed source? You can write a backend wrapper for that shit and hide your prompt logic before it hits LLM api. My guess is that they just didn't beautify source code for public release and did not want to invest in maintaining it as open-source.
OpenAI released their Codex tool open-source. Unless Anthropic makes a bold move to trump that, (opensource always wins), OpenAI will decrease stream of "workflow" user data coming Anthropic's way.
These tools are designed for and must send data to companies that own them. https://github.com/openai/codex/issues/26 OpenAI invests billions into developing intelligence and yet somehow the bottleneck is an API data format - that's why this tool must use OpenAI's proprietary backend only and non other API would do. Think about it. A billions poured into model development, yet an API support cant be developed. Neither company itself nor opensource community is advised (and actively prevented from) "fixing" and API / creating an API wrapper / thinking of a way to achieve this in general. These AI models itself are capable of implementing any API design in seconds, and yet...
Here is even more interesting issue https://github.com/openai/codex/issues/106. You can ask a BS who would bs its way into how it is because of using Responses that depend on B, and B means C, C depends on ..., and now it all makes perfect sense. Or you can ask Elon Musk who would tell you that such logic is an equivalent of 1+1 = fish (common carp). Just smile wide while selling this.
You may try to update these tools and make them work with other LLM providers or even self-hosted ones. Anthropic took down multiple repos on gh that removed login requirement from Claude Code, even though it still was hardcoded to use Anthropic's API. Now ask yourself: why?
The moat game of LLM web UIs is over. Onto building coding tools, let there be no clear winner and we, dear mortals, continue to enjoy cheap technology as a pissing contest between bubble boys keeps unraveling.
Final note: what is ROI for these tools.
The term "Cynic" originates from the Greek word "κυνικός" (kynikos), which means "dog-like".