Openai manuscript is part of a new set of agent coding tools

Last Friday, Openai introduced a new coding system called Codex, designed to perform complex programming tasks from natural language orders. Codex Openai is transferred to a new set of agent coding tools that just began to crystallize.

From the early Copilot of GitHub to contemporary tools such as the indicator and Windsurf, most artificial intelligence assistants work as an exceptionally smart form of automatic completion. Tools generally live in an integrated development environment, and users interact directly with the code created by artificial intelligence. The possibility of setting a task and returning when completing is very far.

But these are new agent coding tools, led by products like Devinand an agencyand openAnd the above mentioned codex, designed to work without users having to see the symbol. The goal is to work like an engineering team manager, set problems through workplace systems such as Asana or Slack and check a solution solution.

For those believers in very artificial intelligence, it is the next logical step in natural automation progress that takes more and more software work.

“Initially, people have just wrote a symbol by pressing every one-key pressure,” explains Killian Lyrett, Prinston researcher and a member of SWE-Agent. “Github Copilot was the first product to provide a real automatic completion, a type of second stage. You are still in the episode completely, but sometimes you can take a shortcut.”

The goal of fake systems is to completely overcome developers environments, instead to present coding agents with a problem and leave them to solve them alone. “We are returning things to the management class,” Lyret says.

It is an ambitious goal, and so far, it has been proven difficult.

After Devin became generally available at the end of 2024, it tied hardness cash From YouTube critics, as well as Criticism more compared to From an early agent in Answer. The general impression was familiar to the old warriors who guide the atmosphere: with many errors, supervision of models takes a lot of work like doing the mission manually. (Although Devin’s Prort was a bit rocky, it did not prevent donations from recognizing the capabilities – in March, it is said that the parent company Devin, AI, is said to. Rapid hundreds of millions of dollars with a rating of $ 4 billion))

Technology supporters are even warning against coding the atmosphere that is not supervised, and seeing new coding factors as strong elements in the process of developing human supervision.

“For the moment, and I would like to say, in the foreseeable future, a person must intervene at the time of reviewing the code to look at the written code,” says Robert Brennan, CEO of all Hands Ai, who maintains open tools. “I have seen many people working in a state of chaos by just a spontaneous breeding for each code written by the agent. It is out of control quickly.”

Halosa is a continuous problem as well. Brennan recalls one incident, when asked about the application programming interface that was released after the OpenHands agent training, the agent manufactured the details on the application programming interface that is appropriate for the description. All Hands Ai says it works on systems to arrest this hallucinations before causing harm, but there is no simple solution.

It can be said that the best measure of progress in programming is Swey-Bench leadersWhere developers can test their models against a set of issues that have not been resolved from the open Gabbab warehouses. OpenHAnds is currently ranked first on the verified leaders panel, 65.8 % of the problem of problems. Openai claims that one of the models that operate Codex, Codex-1 can work better, as he recounted 72.1 % in its announcement-although the result came with some warnings and was not verified independently.

Anxiety among many in the technology industry is that high reference grades do not necessarily translate into the agent’s agent coding. If working programmers can solve three out of every four problems, they will need great supervision of human developers – especially when dealing with complex systems with multiple stages.

Like most artificial intelligence tools, hope is for improvements in basic models at a fixed pace, which ultimately allows the coding systems from growth to reliable developed tools. But finding ways to manage hallucinations and other reliability issues will be very important to get there.

“I think there is little effect of the sound barrier,” says Brennan. “The question is, how much confidence you can move to the agents, so they take more work burden at the end of the day?”