Atlassian has introduced a coordinated system for agentic engineering that connects autonomous AI agents across the software development lifecycle inside Jira. The new capability, called agent loops, allows a backlog item to flow automatically from planning through code generation, testing, and into a reviewed pull request without constant human prompting at every step.
The system relies on what Atlassian calls the Teamwork Graph, a live knowledge structure that maps how teams actually work. It pulls together information from Jira issues, Confluence pages, Bitbucket repositories, Compass components, and historical pull request patterns. When an agent begins work on a ticket, it first consults this graph to understand the surrounding context: which services are affected, how similar changes were implemented before, who owns related code, and what acceptance criteria mattered in past comparable stories. That shared memory reduces the common problem of agents producing code that looks correct in isolation but conflicts with team standards or architecture decisions.
Once the loop starts, a product agent interprets the backlog item and breaks it into granular tasks. A coding agent then generates implementation, drawing directly from the Teamwork Graph to match existing patterns, naming conventions, and error-handling approaches already used in the codebase. The generated code moves to an automated testing agent that writes and runs unit, integration, and contract tests. If failures appear, the loop returns control to the coding agent with specific failure logs and reattempts the fix. Only when all tests pass does the system open a pull request, complete with a detailed explanation of changes, links to related tickets, and suggested reviewers based on ownership data in the graph.
Atlassian demonstrated the loop handling an end-to-end feature addition in a sample microservices application. The entire cycle from backlog item to merged pull request completed with minimal engineer intervention beyond final approval. Early internal testing showed that loops can reduce cycle time on well-scoped tickets by more than half compared with traditional developer workflows, although results vary by complexity and team maturity.
Alongside the agent loops announcement, Atlassian released new developer experience measurement tools integrated into Jira and Compass. These tools track a set of signals that research has connected to actual engineering output: deployment frequency, lead time for changes, change failure rate, and time spent in context switching. The measurements go beyond simple velocity metrics by incorporating qualitative signals such as satisfaction scores from periodic developer surveys and the volume of blocked time logged against tickets.
The DX analysis accompanying the launch examined data from hundreds of organizations using Atlassian platforms. Teams whose members maintained the highest density of connections inside the Teamwork Graph—meaning they actively updated documentation, linked related issues, and kept component ownership information current—shipped approximately 64 percent more features per developer over a six-month period than teams with lower graph density. The correlation held after controlling for team size, domain complexity, and engineering tenure. Atlassian presents the finding as evidence that shared organizational memory becomes even more valuable when AI agents can read and act on that memory at scale.
The company positioned the combined release as a practical step toward what it calls coordinated agentic engineering rather than isolated point tools. Instead of developers copying context into separate AI chat windows, the system keeps every agent operating inside the same information environment that human teams already use. Jira remains the single source of truth: agents create sub-tasks, update status, attach artifacts, and log decisions without requiring users to switch applications.
Security and governance received attention in the design. All agents operate under the same permission model as the human user who triggered the loop. An agent cannot access repositories or services that the triggering engineer could not reach. Administrators can set organization-wide policies that require human review before any generated code reaches production branches. Audit logs capture every decision an agent makes, including which parts of the Teamwork Graph it consulted and why it chose particular implementation approaches.
Early adopters include both established enterprises and smaller product teams. One financial services company reported using loops to handle routine API extensions, freeing senior engineers to focus on complex regulatory requirements that still demand deep human judgment. A gaming studio adopted the system for repetitive gameplay feature scaffolding, noting that the generated code consistently followed the studio’s established architectural patterns because the Teamwork Graph captured those patterns accurately.
Atlassian built the loops on top of its existing Forge platform, which allows the agents to run in a secure, isolated environment close to customer data. The architecture avoids sending sensitive code or proprietary business logic to external large language model providers unless customers explicitly choose to route certain workloads through models hosted by OpenAI, Anthropic, or Google. Most processing happens inside models fine-tuned by Atlassian on anonymized public code and internal best-practice patterns.
The DX measurement tools introduce a new dashboard inside Jira that visualizes four key quadrants: throughput, stability, satisfaction, and cognitive load. Throughput combines deployment frequency with story points completed per engineer. Stability tracks failure rates and mean time to recovery. Satisfaction pulls from short pulse surveys sent through Jira. Cognitive load measures time spent waiting for builds, attending unplanned meetings, and chasing context across tools. The system can segment these metrics by team, product area, or individual contributor, allowing engineering leaders to identify specific bottlenecks.
One notable pattern in the DX data involves the relationship between documentation hygiene and delivery speed. Teams that kept their component catalog in Compass up to date and maintained clear links between Jira tickets and architectural decision records showed both higher agent success rates and better human developer throughput. When the Teamwork Graph contains stale information, agents generate more incorrect assumptions, leading to increased pull request rework. The 64 percent productivity difference appeared most strongly in organizations that treated knowledge maintenance as an ongoing engineering responsibility rather than a documentation afterthought.
Atlassian plans to expand agent capabilities in coming quarters. Future loops will include design agents that can propose user interface changes consistent with existing design systems, documentation agents that update Confluence pages in parallel with code changes, and release agents that coordinate rollout plans across multiple services. Each new agent type will continue to consult the same Teamwork Graph, creating a growing network of specialized workers that share context without duplication.
The company also announced an open beta for custom agent development. Teams can define their own loop patterns using a visual workflow builder inside Jira. A retail company, for example, might create a specialized loop that automatically generates new product catalog endpoints whenever a business analyst adds an item to the merchandising backlog. Because the custom loop references the Teamwork Graph, the generated code automatically includes the correct pricing service integration patterns already used by that organization.
Feedback from the initial pilot programs highlighted both strengths and limitations. Engineers appreciated that the generated pull requests included clear explanations and linked directly to original requirements, making review faster. However, participants noted that highly novel features or those involving significant refactoring still required substantial human guidance. The loops perform best on work that has clear precedents inside the Teamwork Graph. As organizations expand their graph with more historical decisions and architectural patterns, the range of tasks that agents can handle autonomously is expected to grow.
Pricing for the agent loops follows Atlassian’s existing model of per-user subscriptions with additional consumption charges based on the number of loop executions. Organizations already on Premium or Enterprise plans for Jira and Bitbucket receive basic loop functionality at no extra cost during the initial rollout period. Advanced DX analytics and custom agent development carry separate add-on fees.
The launch reflects Atlassian’s broader strategy of embedding intelligence directly into the tools where engineering work already happens. Rather than asking developers to adopt new AI platforms, the company brings agent capabilities into the familiar Jira interface and ties them to the organizational memory already captured in its products. The combination of autonomous loops and quantitative developer experience measurement gives both individual contributors and engineering leaders concrete data about where AI assistance accelerates delivery and where human expertise remains essential.
As more organizations begin experimenting with these coordinated agents, the industry will gather evidence about which types of software work benefit most from autonomous execution and which still require close human oversight. Atlassian’s early data suggests that the quality of shared team knowledge may prove as important as the sophistication of the underlying models. Teams that invest in maintaining accurate, connected information appear positioned to gain the largest advantage as agent loops become a standard part of the development process.
The release marks a concrete milestone in moving AI assistance from individual productivity aids toward coordinated systems that can carry out multi-step engineering workflows. By grounding those workflows in the actual collaboration patterns of each organization, captured in the Teamwork Graph, Atlassian aims to make agentic development feel like an extension of existing team practices rather than a replacement for them. The accompanying DX measurement tools provide organizations with visibility into whether those agents are delivering measurable improvements in both speed and developer satisfaction.
Over the next year, continued refinement of the loops, expansion of available agent types, and deeper integration with the measurement dashboard should give engineering organizations richer options for deciding which work to assign to autonomous systems and which to keep in human hands. The 64 percent productivity correlation observed in the initial analysis offers an early indicator that attention to organizational memory may become a competitive differentiator in an environment where AI agents can read and act on that memory at scale.
from WebProNews https://ift.tt/nec8Gwa
No comments:
Post a Comment