Specialists your team can call on — the same way every time
A custom agent is a ready-made specialist: a planner, a reviewer, a test writer. Each has a focused brief, only the tools it needs, and access to Azure DevOps — so anyone on the team gets the same quality of plan, review or release notes, in GitHub Copilot or Claude Code.
Everyone is writing their own specialist prompts
Each developer asks for a "code review" in their own words and gets a different depth, format and focus. Nothing is comparable.
A general-purpose agent can edit files and run commands even when you only wanted an opinion. That's more risk than a review or plan needs.
People paste work items, acceptance criteria and PR descriptions from Azure DevOps into chat — and the agent still misses linked items and comments.
An agent is a role with a brief, a toolbox and an output format
Each agent is one Markdown file: its purpose, a step-by-step workflow, the tools it may use, and the exact shape of its answer. The same file is used by Copilot, and a Claude Code version is generated from it.
- 1Pick the specialist
Choose ADO Planner, Code Reviewer or another agent from the Copilot agent picker, or ask Claude Code to use it.
- 2It gathers the right context
It reads the work item, pull request or build from Azure DevOps and explores the code — within the tools it was given.
- 3It returns the agreed deliverable
A plan, a findings table with severities, user stories with acceptance criteria, or release notes — the same structure every time.
What teams get
Planning, reviewing and release notes follow the same steps for everyone, so results are comparable across people and teams.
Each agent only gets the tools it needs. Reviewers and planners cannot edit code; the Story Writer asks before creating work items.
Agents read work items, acceptance criteria, pull requests and build logs directly, instead of relying on what someone pasted.
Plans, review findings and release notes arrive in an agreed format that reviewers, testers and product owners can scan quickly.
Security and code reviews run before a human reviewer looks, so people spend their time on design questions instead of routine issues.
One source file per agent. Copilot uses it directly; the Claude Code subagent is generated from it, so both teams get the same behaviour.
The workflow a strong senior engineer would follow is written down once — and every developer can run it.
When an agent misses something, anyone can propose a better brief. Changes are reviewed, versioned and rolled out to every team.
Same request, with and without an agent
Prompt: Plan the implementation for work item 4521.
Without
- You open Azure DevOps, copy the description and paste it into chat
- Acceptance criteria and linked items are missed
- The assistant starts editing code before the approach is agreed
- The plan has a different shape every time
- Risks and open questions are left for the reviewer to find
With ADO Planner
- Reads the work item, acceptance criteria, links and comments itself
- Explores the codebase read-only and asks about gaps first
- Offers approach options with trade-offs
- Returns step-by-step tasks, files to touch, a test plan and risks
- Hands over to implementation only when you agree
Illustrative example based on what the ADO Planner agent instructs. Results depend on the tool, the model and the codebase.
Where each agent fits
| Stage | Agent | What it does |
|---|---|---|
| Plan | Story Writer | Turns ideas, PRDs and bug reports into Azure DevOps stories with acceptance criteria |
| Design | ADO Planner | Turns a work item into a reviewed implementation plan before any code is written |
| Build & test | Test Writer | Writes and runs tests for new or changed code, mapped to acceptance criteria |
| Review | Code Reviewer, Security Reviewer | Severity-ranked findings on a branch or Azure DevOps pull request |
| Release | Release Notes | User-facing notes plus a risk and rollout summary from merged PRs and work items |
How to show it's working
Measure a baseline before rollout, then compare after a few sprints. These are the signals that move first — track them per team, not per person.
| Measure | Why it matters |
|---|---|
| Time from work item "Active" to first pull request | Better plans up front mean less rework and faster starts. |
| Review turnaround and rounds per PR | Routine issues are caught before a human reviewer looks. |
| Stories with clear acceptance criteria | The Story Writer makes testable stories the default. |
| Issues found before merge vs after release | Earlier security and code review should shift findings left. |
| Time to prepare release notes | Release notes become a review task instead of a writing task. |
Useful at every level of the team
Plans, reviews and tests on demand, without writing a long prompt each time.
A consistent first review on every PR, so their review focuses on design.
Well-formed stories in Azure DevOps from rough notes, with nothing created until they confirm.
Draft release notes and a risk summary straight from Azure DevOps.
Pick the agents your team needs
Each item in the catalog has copy buttons and install steps for GitHub Copilot and Claude Code. Claude Code teams can install all 6 agents at once as a plugin.