Coding agents are now available in VS Code and JetBrains IDEs, expanding developer tool choices as the AI agent trend gains mainstream momentum.
GitHub Copilot agent mode became generally available April 4, including a tie-in with Microsoft’s Visual Studio Code integrated development environment (IDE). This week, JetBrains made its Junie coding agent available to users of its IDEs, which include IntelliJ IDEA, PyCharm, WebStorm, PhpStorm, GoLand and RubyMine.
Coding agents were already available from vendors such as Windsurf and Cursor. Atlassian also introduced a coding agent with an update to its Rovo product this month. JetBrains and Atlassian both adjusted pricing to fold AI agents into existing subscriptions, which for JetBrains includes a free tier limited to local AI models.
With AI agents, unlike the previous generation of AI coding assistants that provided inline code snippet suggestions, developers can delegate whole tasks with a natural language prompt such as “Create a web application in Python.”
“It’s a very early form of agentic AI,” said Diego Lo Giudice, an analyst at Forrester Research.
Truly agentic AI networks involve multiple agents working autonomously to carry out multi-step workflows and solve complex problems. However, the wider infrastructure to support agentic AI is still being built with projects such as Model Context Protocol.
For now, these “agentish” systems, as Lo Giudice calls them, will generate more significant portions of code than the previous crop of coding assistants, and some will start connecting tasks in the software development lifecycle.
“That’s what getting more agentic will improve over time — pulling together more complex software development workflows,” he said. “You now have the full process of going from an issue to code, and you can basically submit the pull request and then move on from there.”
Coding agents still need supervision
An initial demo of GitHub Copilot agent mode showed promise for one GitHub Enterprise user, but it still requires a skilled and experienced human to oversee it.
“I gave it a task of refactoring some code, rewriting existing code that works, and it ran for about 10 minutes, and it sort of did what I asked for, and it broke some stuff around it,” said Kyler Middleton, principal software engineer at healthcare tech company Veradigm.
From there, the agent continued to fix what it had broken but introduced new errors, Middleton said.
“It has refactored the code in such a way that it doesn’t even look like my original code, and it keeps introducing problems,” she said. “It is not successfully finishing the task — it’s been running for 90 minutes, and it’s not a very complicated ask.”
Agent mode is more like having a junior engineer writing code for you … and you provide feedback. It’s a totally different experience.
Kyler Middleton Principal software engineer, Veradigm
Middleton said the coding agent concept is also very different from using a coding assistant.
“I use Copilot all the time, but I still feel like I’m the primary author of the code, and I’m just being helped, which is what Copilot aims to do,” she said. “This agent mode is more like having a junior engineer writing code for you … and you provide feedback. It’s a totally different experience.”
With the release of Copilot agent mode to general availability, GitHub also introduced public preview support for Model Context Protocol, which connects coding agents to data sources and secondary tools that ground them in a user’s specific context, boosting the quality of the code they produce.
GitHub is “all in on agent mode and shipping updates daily,” according to a company spokesperson in an emailed statement to Informa TechTarget. “Early data suggests enthusiasm and significant growth in daily usage. There are still many improvements we’re delivering to make agent mode even better, and we look to user feedback to help guide our ongoing investments.”
‘Vibe coding’ trends, but a long way off for enterprises
Generating code has been among the most effective uses for generative AI so far, said Andy Thurai, independent analyst at The Field CTO.
“While the accuracy has improved over the years, it is still not 100% accurate, but based on my conversations with many enterprise developers, it cuts down the coding time tremendously,” Thurai said. “Especially when developers are racing against time during a major incident, to roll out a fixed code quickly and have the systems back up and running, coding agents could be very helpful.”
Andy Thurai
However, there is still the risk that what coding agents produce, deployed in production as is, could add to tech debt over time; enterprises are starting to use code profiling and other validation techniques to reduce this issue, Thurai said.
Coding agents have introduced the concept of vibe coding, in which developers use only natural language to quickly create an application and evaluate the results based on whether the application subjectively looks like they imagined it would. GitHub invoked the trendy term in its blog post about the agent mode rollout.
But while that may be helpful for startups looking to quickly develop a prototype app, enterprises won’t be able to vibe code their way into a scalable, reliable production application — at least so far, Lo Giudice said.
“It will definitely enhance prototyping and innovation in companies, but then somebody is going to have to go and do the rest of the work for turning that into a real, scalable product,” he said. “That’s going to be done by true developers, and not by vibe coders.”
Beth Pariseau, a senior news writer for Informa TechTarget, is an award-winning veteran of IT journalism covering DevOps. Have a tip? Email her or reach out @PariseauTT.
Andy Thurai