Rovo Dev

Rovo Dev is Atlassian's coding agent, available internally on all Jira sites. It works from a Jira work item to a pull request, and the developer reviews the plan and the code at each step. It became generally available in October 2025.

Company: Atlassian
Approach type: Agent
Work: Coding, Code review
Human involvement: Human in loop
Invocation: Interactive
Interfaces: Jira
Deployment stage: Scaled
Evidence strength: Detailed primary
Entry reviewed: 2026-09-16

Copy as MD Ask ChatGPT

How it works

Jira work item through the four HULA phases to a reviewed pull request.

Set the context

The user provides the repository and the work-item description the agent needs^{[Source 1]}

Generate the plan

The agent generates a plan that the user can review and refine before it proceeds^{[Source 1]}

Generate the code

The agent writes the code, and the human can review it and request changes^{[Source 1]}

Raise the pull request

The agent raises the PR, and the user reviews it and merges it^{[Source 1]}

Where people stay involved

People review, approve, or guide the agent’s work here.

Reported: The user reviews the plan, the code, and the PR; the code-reviewer side leaves accept-or-decline decisions to the human reviewer.

Supervision level
Work-product review
A person reviews the draft or implementation.
Level 3 of 5
- Jira issue → reviewed pull request^{[Source 1]}

How supervision levels work

Implementation details

Model

Rovo Dev Code Reviewer runs Anthropic Claude 3.5 Sonnet, with a gpt-4o-mini judge for factual checks^{[Source 2]}Documented for the code reviewer; the Jira coding agent model is not named.

Harness

Built on the HULA framework, which runs set context, generate plan, generate code, and raise PR^{[Source 1]}

Sandbox

Not reported

Tool access

Not reportedRepository selection is named as user-provided context, not as a documented tool-access mechanism.

Knowledge

Not reportedNo separate knowledge store is documented; similar work items are recorded under context management.

Context management

Top-50 similar resolved Jira work items and their merged PR diffs attach as additional run context^{[Source 1]}

Credentials

Not reportedPrivacy and data security are named as design concerns, not as a documented credential mechanism.

Interfaces

Jira for the coding agent; the code reviewer integrates with Bitbucket (source-2 lines 41, 83), which the interface vocabulary does not list.

jira

^{[Source 1, Source 2]}

Validation and failure handling

Filter review comments through two checks

A gpt-4o-mini judge drops factually incorrect comments and a fine-tuned ModernBERT drops unactionable ones before posting^{[Source 2]}

Reported observations

These results come from the sources. Notes below each result explain what it covers and any reported limits.

Headline claim

Dogfooded across 1,900+ repositories with a 50,000+ comment internal dataset^{[Source 2]}

Reported by: Atlassian
Scope: Internal code-review dogfooding repositories and Rovo Dev-generated classifier training comments

The source does not report the denominator of this figure.

Observed in 2026

About this result

Category: Implementation scale
Evidence type: Reported measurement
What it covers: Code-review dogfooding repositories and the classifier training dataset

Key observation

Median PR cycle time reduced by 30.8% over the year-long evaluation^{[Source 2]}

Reported by: Atlassian
Scope: Median PR cycle time in the code-reviewer dogfooding evaluation
Method: Year-long online evaluation across 1,900+ internal repositories

The source does not report the denominator of this figure.

Observed in 2026

About this result

Category: Effectiveness
Evidence type: Reported measurement
What it covers: Median PR cycle time in the code-reviewer evaluation

Key observation

Human-written review comments reduced by 35.6%^{[Source 2]}

Reported by: Atlassian
Scope: Human-written review comments per PR in the code-reviewer dogfooding evaluation
Method: Year-long online evaluation across 1,900+ internal repositories

The source does not report the denominator of this figure.

Observed in 2026

About this result

Category: Effectiveness
Evidence type: Reported measurement
What it covers: Human-written review comments per PR in the code-reviewer evaluation

Lessons

Not reported: The HULA abstract references lessons learned but the preserved capture states none; the blogs give design philosophy, not evaluated lessons.

Back to agents

Improve this record

Give your coding agent the context and instructions to prepare a source-backed update.

Suggest on GitHub

Paste the prompt into your coding agent, then describe the correction or add a public source.

Preview prompt

Help me update the Internal Agents Map record for Rovo Dev at Atlassian.

Repository: https://github.com/steel-experiments/internal-agents-map
Record: data/agents/atlassian-rovo-dev.yaml
Published record: https://internal-agents.com/agents/atlassian-rovo-dev
Readable context: https://internal-agents.com/agents/atlassian-rovo-dev.md

Use the correction or public source I provide in this conversation. If I have not supplied either, ask what I want changed before editing.

Use an existing checkout of this repository, or clone it if needed. Read the current record from the repository. Follow AGENTS.md, CONTRIBUTING.md, and the relevant instructions in data/schema.md.

Verify that the evidence supports the requested change. Preserve uncertainty and distinguish reported facts from catalog judgments. Follow the repository process for capturing new sources. If evidence is insufficient, explain the gap instead of inventing details.

Make a focused change on a branch. Regenerate derived files and run the required checks. Do not edit generated content by hand.

Show me the change, supporting evidence, and check results, including any checks you could not run. Ask whether to submit a pull request unless I already requested one. Do not merge the pull request. If you cannot edit the repository, prepare a source-backed issue draft.

Sources and research details

Citations link to the original publisher. Each source also keeps a preserved copy in the repository, so a changed or removed page stays checkable.

Improving the coding agent experiencehttps://www.atlassian.com/blog/atlassian-engineering/improving-coding-agent-experienceEngineering blog · First party · Last source verification: 2026-08-31
Preserved copy
Developer productivity improved with Rovo Devhttps://www.atlassian.com/blog/atlassian-engineering/developer-productivity-improved-with-rovo-devEngineering blog · First party · Last source verification: 2026-08-31
Preserved copy
HULA: Human-in-the-loop software development agents (arXiv 2411.12924)https://arxiv.org/abs/2411.12924Paper · First party · Last source verification: 2026-08-31
Preserved copy

Question coverage and scope

purpose: Reported
workflow: Reported
human involvement: Reported: The user reviews the plan, the code, and the PR; the code-reviewer side leaves accept-or-decline decisions to the human reviewer.
implementation: Reported
validation: Reported: The two quality gates belong to the code-reviewer side of this record, not to the Jira coding workflow.
observations: Reported: The dogfooding and evaluation figures describe the code-reviewer side; the Jira coding agent has no measured outcome in the captures.
lessons: Not reported: The HULA abstract references lessons learned but the preserved capture states none; the blogs give design philosophy, not evaluated lessons.

Research details for every claim on this page

Summary
Statement type
Fact
Provenance
Reported
Confidence
Not assessed
Confidence reason
No explicit confidence assessment is recorded for this claim.
Headline claim
Statement type
Metric
Provenance
Reported
Confidence
Medium
Confidence reason
Atlassian reported the dogfooding scale in its own engineering blog without independent verification.
Reported by
Atlassian
Scope
Internal code-review dogfooding repositories and Rovo Dev-generated classifier training comments
Denominator
Not reported
Method
Not reported
Observation date
2026
- SupportsDeveloper productivity improved with Rovo DevPreserved content.md, lines 66, 87
Harness
Statement type
Fact
Provenance
Reported
Confidence
Not assessed
Confidence reason
No explicit confidence assessment is recorded for this claim.
- SupportsImproving the coding agent experience
Model
Statement type
Fact
Provenance
Reported
Confidence
Not assessed
Confidence reason
No explicit confidence assessment is recorded for this claim.
- SupportsDeveloper productivity improved with Rovo DevPreserved content.md, lines 41, 60
Context management
Statement type
Fact
Provenance
Reported
Confidence
Not assessed
Confidence reason
No explicit confidence assessment is recorded for this claim.
- SupportsImproving the coding agent experiencePreserved content.md, lines 67–74
Interfaces
Statement type
Fact
Provenance
Reported
Confidence
Not assessed
Confidence reason
No explicit confidence assessment is recorded for this claim.
- SupportsImproving the coding agent experience
- SupportsDeveloper productivity improved with Rovo DevPreserved content.md, lines 41, 83
Set the context
Statement type
Fact
Provenance
Reported
Confidence
Not assessed
Confidence reason
No explicit confidence assessment is recorded for this claim.
- SupportsImproving the coding agent experiencePreserved content.md, line 46
Generate the plan
Statement type
Fact
Provenance
Reported
Confidence
Not assessed
Confidence reason
No explicit confidence assessment is recorded for this claim.
- SupportsImproving the coding agent experiencePreserved content.md, line 47
Generate the code
Statement type
Fact
Provenance
Reported
Confidence
Not assessed
Confidence reason
No explicit confidence assessment is recorded for this claim.
- SupportsImproving the coding agent experiencePreserved content.md, line 48
Raise the pull request
Statement type
Fact
Provenance
Reported
Confidence
Not assessed
Confidence reason
No explicit confidence assessment is recorded for this claim.
- SupportsImproving the coding agent experiencePreserved content.md, line 49
Filter review comments through two checks
Statement type
Fact
Provenance
Reported
Confidence
Not assessed
Confidence reason
No explicit confidence assessment is recorded for this claim.
- SupportsDeveloper productivity improved with Rovo DevPreserved content.md, lines 60–70
Key observation
Dogfooded across 1,900+ repositories over more than a year
Duplicate representation of Headline claim. Component of the compound headline: the same 1,900-plus repository count over the same year-long evaluation.
Statement type
Metric
Provenance
Reported
Confidence
Medium
Confidence reason
Atlassian reported the dogfooding scale in its own engineering blog.
Reported by
Atlassian
Scope
Internal code-review evaluation across 1,900+ repositories spanning over a year
Denominator
Not reported
Method
Year-long online evaluation across internal repositories
Observation date
Not reported
- SupportsDeveloper productivity improved with Rovo DevPreserved content.md, lines 87
Key observation
Trained on a proprietary internal dogfooding dataset of 50,000+ Rovo Dev comments
Duplicate representation of Headline claim. Component of the compound headline: the same 50,000-plus comment dataset.
Statement type
Metric
Provenance
Reported
Confidence
Medium
Confidence reason
Atlassian reported the dataset size in its own engineering blog.
Reported by
Atlassian
Scope
ModernBERT classifier training dataset of internally sourced Rovo Dev-generated comments
Denominator
Not reported
Method
Internal dogfooding comments labeled by whether they led to a code resolution
Observation date
Not reported
- SupportsDeveloper productivity improved with Rovo DevPreserved content.md, lines 66
Key observation
Statement type
Metric
Provenance
Reported
Confidence
Medium
Confidence reason
Atlassian reported the reduction from a year-long internal evaluation accepted at ICSE 2026; no independent replication exists.
Reported by
Atlassian
Scope
Median PR cycle time in the code-reviewer dogfooding evaluation
Denominator
Not reported
Method
Year-long online evaluation across 1,900+ internal repositories
Observation date
2026
- SupportsDeveloper productivity improved with Rovo DevPreserved content.md, lines 16, 87–91
Key observation
Statement type
Metric
Provenance
Reported
Confidence
Medium
Confidence reason
Atlassian reported the comment reduction from the same internal evaluation; no independent replication exists.
Reported by
Atlassian
Scope
Human-written review comments per PR in the code-reviewer dogfooding evaluation
Denominator
Not reported
Method
Year-long online evaluation across 1,900+ internal repositories
Observation date
2026
- SupportsDeveloper productivity improved with Rovo DevPreserved content.md, lines 16, 87–89
Operating model assessment
Statement type
Inference
Provenance
Catalog judgment
Confidence
Medium
Confidence reason
The HULA framework and engineering blog describe a human-in-the-loop cycle that ends in a reviewed pull request.
Observation date
2024-11
- SupportsImproving the coding agent experience