Atlassian’s $1.2B Case for Agent Control

In February 2026, workflow-software maker Atlassian was the Nasdaq 100’s worst performer, down more than 45% as investors feared AI agents would hollow out its products. Six months later, its Q4 results sent the shares up roughly 30%. Which system decides whether an agent may act?

As model vendors make code generation more interchangeable, the durable enterprise prize shifts to the control plane around AI agents: the system that supplies trusted context, grants execution rights, routes approval, governs deployment and preserves an audit record.

Engineering organizations do not accept a patch because it compiles. They bind requests to issues, repositories to owners, changes to tests, tests to approvals, deployments to environments and incidents to response logs. Human developers navigate those bindings through organizational memory. Agents need software to make them explicit.

The coding environment can now swap its model

Apple made the supply shift unusually clear. Xcode 26.3 supports both Anthropic’s Claude Agent and OpenAI’s Codex, alongside the Model Context Protocol. Apple owns the development environment, yet it chose a multi-provider agent layer rather than requiring one proprietary coding intelligence. The IDE can remain stable while the model changes underneath it.

Anthropic has pushed in the other direction, increasing the amount of work a single agent system can coordinate. Claude Code’s dynamic workflows can run hundreds of subagents in parallel on engineering jobs such as framework migrations. The Allen Institute for AI widened supply from another side with open-source coding-agent models designed to adapt to private codebases. These independent moves point toward the same architecture: falling generation costs and replaceable interfaces.

Hundreds of parallel subagents multiply the need for live information and external tools. A model can compose syntax from its training and prompt, but an engineering agent must retrieve the current issue, inspect the repository, query build results, invoke tools and respond to feedback. Each connection introduces an owner, a permission boundary, a rate limit and a possible stale state.

A coding model can attempt the first row. Enterprise software earns its place by keeping the other four current while many models and tools pass through them.

A work graph gives an agent somewhere to stand

Atlassian began putting models inside its existing work systems before coding agents became the organizing idea. Atlassian Intelligence arrived in Jira and Confluence Cloud in 2023 using Atlassian’s models and OpenAI’s. In 2024, Rovo added contextual search, insights and custom agents intended to handle tasks and “complete projects.”

Jira and Confluence give an agent different forms of state. A Jira issue can identify the requested change, its status and its relationship to other work. Confluence can hold the operating knowledge around that request. Automation can move the issue or invoke another system after a condition changes. The agent needs a reliable account of where the work stands and what action follows.

OpenAI made the control-plane contest explicit in 2026. Its Symphony specification turns a project-management board such as Linear into a control plane for coding agents. The board supplies assignments, state transitions and orchestration around agents that may come from elsewhere.

OpenAI’s choice of Linear also prevents Atlassian from claiming the category by definition. Any project system can attempt this role if it holds current work state and can connect that state to tools. To serve as context, a ticket database must reflect current requests. To serve as control, a board needs identity, policy and deployment hooks. Agents expose stale records and missing connections with admirable lack of tact.

Atlassian built the workflow stack before coding agents

Atlassian entered the agent era with a workflow portfolio assembled across planning, automation, delivery and incident response. Each product supplied a block that autonomous work now requires.

• 2018: Atlassian bought OpsGenie for $295 million, extending from planning and tickets into operational incident response.
• 2019: Atlassian acquired AgileCraft for about $166 million to add strategic-project and workstream planning.
• 2019: Atlassian acquired Code Barrel, whose Automation for Jira handled routine processes at companies including Visa, Cisco and Airbnb.
• 2020: Atlassian launched native no-code automation across Jira Cloud products, with integrations including Slack and Microsoft Teams.
• 2023: Atlassian placed AI assistance inside Jira and Confluence Cloud.
• 2024: Rovo combined contextual retrieval with custom agents intended to act on work.

Together, these products give Atlassian a credible base for workflow-native AI. It can connect a requested change to planning, documentation, automation and operational response without asking customers to build an entirely new work system around the model.

Atlassian still has to turn adjacent products into one governed environment. Jira may know who owns an issue; an incident system may know which service is failing; a delivery system may know which build reached production. An agent needs those systems to agree before it acts, not during the retrospective.

Atlassian can use its enterprise customers, administration tools and cross-team visibility to finance the control layer. Those same requirements make a rapid rewrite harder than it is for a new developer tool. Atlassian must convert its installed workflow into interoperable infrastructure without breaking the processes that made customers buy it.

Customers kept buying while investors sold

Atlassian reported Q4 revenue of $1.77 billion, up 28% year over year and above the $1.66 billion estimate. The company forecast Q1 revenue above estimates and returned to profit.

Customers kept buying Atlassian’s substrate as agents became more capable. The results do not isolate how much demand came from Rovo, the Teamwork Graph, enterprise sales or the underlying product suite. Mike Cannon-Brookes’ planned $250 million share purchase expressed conviction after the rebound; it did not provide product-level attribution.

In Q3, Atlassian’s revenue rose 32% to $1.79 billion while its net loss widened. The company also cut roughly 1,600 jobs, or 10% of its workforce, to fund AI and enterprise-sales investment, with expected charges of $225 million to $236 million. Atlassian is rebuilding its operating model while selling the new one.

By February 2026, Atlassian’s founders had lost roughly $7.2 billion in wealth as TEAM sank. The Q4 rebound challenged the idea that agents simply erase workflow software, but one strong quarter cannot establish durable AI-era margins.

Review gates make autonomous work governable

An agent capable of coordinating hundreds of subagents changes the unit of oversight. A reviewer cannot supervise every intermediate token or tool call. The organization instead defines boundaries around identity, accessible systems, permitted actions, test requirements, approval thresholds and deployment environments.

Enterprises create orchestrated autonomy by letting the model choose among actions inside a governed architecture. A human approval gate catches defective code and assigns responsibility at the point where generated work becomes an organizational act.

Enterprises therefore need deployment accountability as product infrastructure. They must know which agent changed a file, which model and tools it used, which policy allowed the action, who approved the result and what reached production. The record matters even when the code works, because accountability cannot depend on failure being obvious.

Verifiable agent control and agent-execution governance belong in the workflow layer. Repository permissions, sandboxing, approval gates and action logs govern what an agent can make real. Model safety bounds what the system may attempt; workflow governance bounds what the enterprise lets it complete.

Every adjacent SaaS category can bid for control

Atlassian faces competitors from four directions. Project-management challengers can turn their boards into agent state machines. Linear raised an $82 million Series C at a $1.25 billion valuation before OpenAI named its board as Symphony’s example control plane. Its advantage would come from becoming the cleanest place to assign and supervise model-driven work.

Developer platforms own other valuable boundaries. Microsoft and GitHub can connect repositories, developer identity, VS Code, project rules and delivery workflows. Cursor’s Automations can trigger agents from a codebase change, a Slack message or a timer, moving orchestration out of the chat box and into event-driven processes.

Model companies can build downward from the runtime. OpenAI’s Symphony specification and Anthropic’s integrations let their agents reach project systems and external tools without owning the original record. Open interfaces lower their cost of entry, although those companies still depend on customers to grant access to enterprise context and actions.

ServiceNow can build sideways from its enterprise workflow footprint. It reported Q2 subscription revenue of $3.88 billion, up 24.5% year over year, giving it another base from which to govern AI-mediated work beyond software development.

In 2025, Microsoft, OpenAI, Salesforce and others were still marketing incompatible meanings of “agent”. Vendors that price an agent as a mystical digital employee weaken their case with confused customers. Enterprises can compare products more cleanly when vendors expose the concrete unit—a tool call, an approved task, a deployed change or a resolved incident.

Autonomous work breaks the seat as a complete price

Per-user subscriptions assume that human access approximates product consumption. Agents disturb that relationship. One employee can launch many agents, one agent can perform work across several systems, and background automations can continue without another human seat entering the application.

Software vendors have already started adjusting. By the end of 2025, 79 of 500 tracked software companies had adopted usage-based AI fees, more than twice the 2024 level. Usage pricing lets vendors charge for expanding machine activity, but a token or invocation meter measures consumption rather than business value.

Workflow providers can meter richer events because they see the lifecycle around the action. A project system can observe assignment and completion. A repository platform can observe a merged change. A delivery system can observe a deployment. An incident platform can observe resolution and recurrence. The provider that joins those events can price controlled work against an observable outcome.

Customers can use open protocols to replace the model, move context through standard interfaces and route actions across several tools. That portability prevents any single IDE or project board from inheriting monopoly power. Atlassian, Linear, Microsoft, OpenAI, Anthropic and ServiceNow must earn control by keeping context accurate, integrations broad and policy enforcement credible.

Atlassian’s reversal shows why the expensive question begins after the patch exists. The repository holds code; the work system decides whether an agent may merge it, deploy it or close the resulting incident. That permission is the product.