2025: Year Of The Agents To Write All Of Our Codex?

"2025 is the year of agents!" - Greg Brockman

I can’t quiiiiite tell if Greg is joking here in that so many have been using this as a rally cry, but it’s definitely true… at least from the perspective of where many are focused.

We are moving along and pushing up in the chart from attended ⇒ autonomous.

OpenAI went on to re-introduce something called Codex. There was their first code model waaaaaay back in history, their local CLI which mimics Claude Code, the new model codex-1 , and now we have the Codex Agent:

This builds on the CLI in that it is a remote agent… a member of the team that you can farm issues too, interact with, and has some key components:

• Isolation: It’s not running on your local machine… that’s the remote part. It is running in OpenAI’s cloud, which can scale as well as not monkeying with your machine. What level of isolation is being used to keep your world safe from anyone else's?

• Integrations: You will want to setup all of the environment variables! It’s acting as another entity and thus needs ways to talk to your systems safely and securely. Which leads you to thinking… which systems will you trust a remote agent with?

• Scale: Being able to fire off a series of threads where Agents are able to explore and do work, even when you close the laptop lid, is liberating and if the Agents are able to do a good job… incredibly profitable. Matt Holden, a phenomenal current solopreneur wrote about his experience with Codex, and how it’s changing his view of AI coding.

To be successful, you will need to invest in some things up front. It can feel a lil like eating vegetables, but think about how good that pina colada will taste when you sip it at the beach as you watch The Robots building for you.

You get what you put in, so… put in the work to setup tests… including integration tests (not just unit). You know that you should have been doing this well all along, because humans are known to make mistakes, so this is just giving the extra push. The more you invest here, the better the results are:

• The agents can read the tests which gives them a nice TDD path and will result in much better context and awareness (if you write the tests, that’s the best… if you don’t, they can write them… from your nice spec ofc!)

• The agents can run the tests and try to auto fix things

• You will be able to watch over what’s happening and will have a lot more trust behind it all

As we mentioned last week, just be careful that a) the tests are good, and b) the Agents aren’t trying to “just get the damn green success!”.

But what about that CLI?

OpenAI released the Codex CLI in April. That’s really recent. The CLI was in response to the success that Claude Code was having with their local agent CLI, which was released way back in …. February. How time flies in AI land!

Let’s take a look at both:

By making Claude Code closed source, it gave strategic room for OpenAI to jump on its success and quickly ship an open source version and do an end run. The Codex CLI is also open in another very important way… it supports multiple models and providers. Importantly it defaults to OpenAI, so if they can keep devs using it for other models vs forks and other CLIs, they can stay in the catbird seat… and get data on usage.

Now, Claude Code is a closed tool just for their model, which is now competing with an open source version, backed by a leading frontier lab, and with the ability to use alllll of the models.

These are all trade offs. For example, by only supporting their own models, Anthropic can deeply integrate and not worry about how to make things work across multiple models.

A look at GitHub

Let’s take a peek at GitHub and see if there has been anything different between the two there:

We all know that GitHub stats contain much vanity and don’t tell a full story, and we really want to know more about raw adoption, but there are some interesting things here:

• OpenAI Codex has 23.2k stars vs. Claude Code’s 8.7k, in a much shorter time to boot (31 vs. 84 days)

• OpenAI Codex has 109 contributors compared to Claude Code's 11 (which are all internal). How meaty are they? We know that we can’t bank on major contributions, but it is nice to see that they can happen, especially with a JS/TS type toolchain which is easy to jump into.

How will the CLI and the remote Agent come together?

With the new Code CLI and agent launching and iterating, how do they fit? Greg makes it clear that he expects a level of convergence and integration:

Time will tell!

ps. We like it when developer services share their prompts, so it was great to see Codex sharing theirs!