AI coding agents work—if teams change how they work
AI coding agents are developer productivity tools that automate or accelerate coding tasks such as generating patches, drafting pull requests, and running commands, and they are increasingly wired into workflows through pull request automation, GitHub Copilot integration, and chat-based interfaces so that developers can move from idea to merged code with far fewer manual steps. A large-scale study at Microsoft’s engineering org found that developers who used command-line AI coding agents—Anthropic’s Claude Code and GitHub Copilot CLI—merged roughly 24% more pull requests over a four‑month period than they otherwise would have. The paper measured a +24.0% increase in merged pull requests per engineer per day, with a likely range of +14.5% to +33.7%, and the gain did not fade during the four‑month study window.
The key detail: this lift appeared only when people used the tools regularly; handing out licenses alone did nothing. Engineers who used the agents five or more days per week saw more than a 50% boost, compared with roughly 15% for those who used them three days a week. In other words, AI coding agents are not magic; they are habits. Teams that treat them as background noise will plateau around the industry‑wide 10–15% productivity bump that Atlassian’s longitudinal work with its DX platform has observed, while teams that reorganize around them can see much larger gains.
More pull requests, more pressure: the new review bottleneck
The headline productivity story hides a harsher truth: AI coding agents move the bottleneck rather than erase it. A separate enterprise study that tracked 802 developers and 196,212 pull requests from January 2024 through April 2026 saw per‑developer throughput more than double, from 21.2 to 44.3 merged pull requests per active developer, after a formal “2x mandate” from the CTO. That sounds like victory until you look at review capacity. As AI‑authored pull requests increased, the share of pull requests receiving at least one human review fell from 89% to 68%, while automated AI review coverage shot from roughly 19% to 84%, and the workload per reviewer roughly doubled.
This is the uncomfortable trade: AI coding agents give teams more output than their reviewers can thoughtfully handle. Across 567 Claude Code pull requests in 157 open‑source projects, 83.8% were eventually merged, but only 54.9% went in without further changes and the rest needed human fixes, especially for bug repairs, documentation, and project‑specific standards. The practical lesson is blunt and quotable: “Before scaling AI coding agents, organizations should track four things: adoption by team, how often each developer uses the tools, how much legacy code is in scope, and whether reviewers can handle the extra pull‑request volume”.
Jira’s AI agents show how workflows—not IDEs—are becoming the home of coding
One company is betting that the future of AI‑assisted development lives inside the work tracker, not the editor. Atlassian has launched a batch of AI‑centric features to pull Jira deeper into the software development lifecycle, including a built‑in Jira Coding Agent that turns work items into pull requests without anyone setting up a local environment. Teams can assign Jira issues directly to Claude Code, Cursor, or GitHub Copilot, with OpenAI’s Codex on the way. In practical terms, today’s updates aim to move more of the entire workflow into Jira: plan with the AI, publish the spec to Confluence, then break it into technical work items that route back into Jira agents.
This is not about yet another shiny developer productivity tool; it is about removing the constant context switching that developers hate. Jira for Slack lets you file work items by asking @Jira inside a thread, while the Teamwork Graph context layer spans Jira, Confluence, Slack, GitHub, and Jira Product Discovery. That web of integrations, including tight GitHub Copilot integration, is quickly becoming the default expectation for AI‑assisted development. Atlassian says an internal study of 6,000 engineers using these capabilities produced a 44% boost in agent task completion efficiency, a 48% drop in token consumption, a 36% reduction in PR cycle time, and 51% of routine code vulnerabilities resolved autonomously and queued for developer review.
Sustained gains demand discipline, not blind automation
The through‑line across these experiments is discipline. AI coding agents can double throughput in newer repositories while leaving legacy systems almost untouched, because they fit greenfield work better than tangled old code. Atlassian’s own longitudinal data with DX found that developer productivity gains from AI topped out around 15% for many organizations. That ceiling is not a technical limit; it is a management one. If teams reward raw pull request counts without guarding review quality, they will drift toward shallow, fragmented changes and rely on AI‑only review to keep up.
There is a longer arc here. Atlassian’s head of engineering for DevAI argues that “a lot of things will eventually be fully automated,” while noting that today “the bottleneck is human” and that long code will remain hard to review. At the same time, planning horizons are shrinking; Atlassian itself avoids roadmaps beyond a quarter, with six months as the outer limit and everything past that “a guess”. The best bet is not to chase a mythical 2x forever but to build stable habits: track real use of AI coding agents, design strong pull request automation and review policies, and invest in reviewers as much as in writers. The teams that treat AI coding agents as a new colleague rather than a toy will be the ones still ahead when the hype cycle moves on.
