Title: The methodology: Agent-Driven Development | Glideship
Source: https://glideship.ai/methodology/
Summary: How Agent-Driven Development works: across nine phases, AI agents draft specs, designs and tests while people review and decide. Plus the manifesto and FAQs.

How it works

# A new way to build software

From a Slack message to production. Agents handle triage, specs, estimates and retrospectives, and people review and decide. No sprints, no story points, no standups.

The problem

## Most of the week isn't code.

Specs, status updates, estimates and meetings fill the rest. That is the work Glideship hands to agents, so people keep the decisions and get the time back.

16%

is how much of their time developers spend coding, Atlassian reports. The rest of the week goes on everything around the code. — Source[Atlassian, State of Developer Experience 2025](https://www.atlassian.com/blog/developer/developer-experience-report-2025)

Industry research from the sources linked above. These are not Glideship results.

Watch one BUD

## From idea to close, in two minutes.

Follow one feature through the phases below, from the first draft to the closing retrospective.

[Watch with sound 2:07](https://www.youtube.com/watch?v=MfvfDH2RrvE)

One BUD from idea to close — Follow a single feature through every stage of its life.

**Read the transcript**

In Glideship, every feature lives in one document called a BUD, a Business Understanding Document. This board has a column for each stage, from the first request to closed. Each card shows the BUD number, priority, complexity, the assignee and any open bugs. You can filter by title, assignee, priority or team.

Daniel creates a new BUD with a title, a short description and a priority. Under advanced settings he chooses which phases the agents should draft for him. Here he turns them off, so the team writes this one by hand. Here is one further along. The header shows the stage, the assignee, and the date it is likely to ship.

The tabs follow the lifecycle. Requirements, design wireframes, the tech spec, implementation to-dos with their pull requests, code review, and the test plan. It is all one document. Each section keeps its history, so you can compare versions and restore an earlier one.

Below the tabs is the go-live forecast. Instead of story points, it gives a range of dates, from an optimistic one to one that is safe to promise. Moving forward is one button. Design, then tech architecture, development, code review, testing, UAT and prod. The page opens the tab that matches the new stage.

Once the work is live in production, the BUD is closed. The people who shipped it get the credit, and the BUD can feed a short retrospective. One document, from the first request to the last lesson.

The flow

## Nine phases. One BUD.

Every feature follows the same path. At each step an agent drafts the work and a person reviews it, changes it and moves it on.

Phase 1 of 9

### A request comes in

Someone asks for something in Slack. The Triage Agent reads it, checks for duplicates and runs a short interview about business impact, who it is for, timing and dependencies.

What agents do

Reads the request, looks for duplicates, sizes the work from the code, and suggests a priority to the product manager.

What people do

Shares the idea and answers the questions. A product manager or the org owner reviews the triage, adjusts it and moves it on to a BUD.

Phase 2 of 9

### Writing the BUD

The BUD becomes the one place for everything: spec, tech spec, test plan and acceptance criteria. The BUD Agent checks your team's rules and earlier work, then drafts every section. The product manager reviews it, sharpens it and moves it to Design when happy.

What agents do

Checks your rules and past work, then writes the overview, goals, user stories, requirements, acceptance criteria, what is out of scope, dependencies and risks. Creates the BUD folder in your repo.

What people do

Reviews the draft spec, sharpens the requirements and moves it to Design when ready.

Phase 3 of 9

### Design

Once the BUD reaches Design, the agents work out what needs designing and draft wireframes. The designer reviews them, edits in the tools they like and moves on to Tech Architecture when happy.

What agents do

Scopes the design work, drafts HTML wireframes in your own design system, suggests patterns, and picks up Figma review notes through MCP.

What people do

Reviews the wireframes, edits in the tools they like and moves on to Tech Architecture when ready.

Phase 4 of 9

### Tech plan and build

After Design, the Tech Plan Agent writes the build plan and the tech lead reviews it. The Smart Assignment Agent suggests the best developer based on skills and workload. If your team has a manager, they confirm the choice. Otherwise the developer is assigned straight away. AI then builds the feature with full access to the code.

What agents do

Writes the tech plan, suggests the best developer through the Smart Assignment Agent, and builds the feature to your team's standards.

What people do

Reviews the tech plan and starts development. A manager reviews the assignment, if your team has one.

Phase 5 of 9

### The test plan

Once the code is written, AI drafts a full test plan: automated tests (unit, integration, end-to-end, performance and security) and manual checks (UAT scenarios, edge cases and exploratory guides). Every test ties back to the BUD's acceptance criteria.

What agents do

Writes unit, integration, end-to-end, performance and security tests, plus manual UAT scenarios and exploratory test guides.

What people do

Reviews the test plan and adds the edge cases only your team would know.

Phase 6 of 9

### QA takes over

QA now owns the BUD. They approve the automated test plan or adjust it, run the manual checks, record proof and sign off. Then the BUD moves to UAT.

What agents do

Presents the test plan for QA to review, tracks the manual checks and collects the proof.

What people do

Approves or adjusts the test plan, runs the manual checks, records proof and signs off for UAT.

Phase 7 of 9

### UAT and release

After QA signs off, the BUD goes through UAT (user acceptance testing) and then to production. The Status Agent notices merged pull requests and works out the stage from the target branch. Everyone who needs to know is told automatically.

What agents do

Spots merged pull requests, updates the BUD status, tells stakeholders and tracks the release.

What people do

Checks the feature in the UAT environment and gives the go or no-go for production.

Phase 8 of 9

### The BUD becomes a Feature

Once it is live, the BUD graduates to a Feature. It leaves the active board and joins the feature registry, a lasting record of what was built, why and how. The BUD's journey is complete.

What agents do

Files the BUD as a Feature in the registry, updates the knowledge base and starts the learning step.

What people do

Confirms the release went well and checks it in production.

Phase 9 of 9

### Learning and skills

The Learning Agent measures how long the work took, compares that with the estimate, finds patterns across similar features, writes the retrospective and stores the lessons so they can be searched later. The Skill Agent refreshes developer profiles every day.

What agents do

Measures cycle time (how long the work really took), writes the retrospective, warns when only one person knows a module, and suggests future assignments by skill and workload.

What people do

Reviews the insights, confirms the lessons and looks after your team's rules.

The loop

### When a bug comes back

Bugs come from outside: production monitoring, user reports or support tickets. When a bug matches an existing Feature, that Feature reopens and the flow starts again from triage, with the bug labelled along the way.

What agents do

Matches the bug to its Feature by meaning, labels it "missed requirement" or "implementation bug", reopens the Feature and starts triage.

What people do

Reports the bug, checks the label and sets the priority of the fix.

Who does what

## Agents do the process work. People decide.

An agent never ships on its own say-so. Each draft lands in front of a person, who can approve it, change it or send it back.

### Agents handle

- Sorting new requests and spotting duplicates

- Drafting BUDs from your real code

- Scoping design work and writing the tech plan

- Writing test cases, automated and manual

- Linking bugs to BUDs and watching bug counts

- Tracking status and keeping stakeholders posted

- Spotting patterns and writing retrospectives

- Mapping skills and suggesting who takes what

- Keeping knowledge in sync with the code

- Delivery dates with real odds (AI-PERT + Monte Carlo)

### People handle

- Reviewing and approving every phase

- Code review and architecture choices

- Visual design, in the tools they like

- Business trade-offs and priorities

- Quality checks and UAT sign-off

- Reviewing or overriding reassignments

- Curating knowledge and team rules

Why people stay in charge

46%

of developers distrust the accuracy of AI tools, more than the 33% who trust it. Only 3% highly trust the output. — Source[Stack Overflow Developer Survey 2025](https://survey.stackoverflow.co/2025/ai)· 33,244 responses

The manifesto

## Eight trades we made.

1. 01
  AI first draftsoverstaring at a blank page

2. 02
  Delivery forecastsoverstory points and planning poker

3. 03
  Learning as you gooverpost-mortems after the damage

4. 04
  Human decisionsoverhuman busywork

5. 05
  Living knowledgeoverstale wiki pages

6. 06
  One BUD per featureoverscattered tickets and docs

7. 07
  Skills that grow with the teamoverfixed role assignments

8. 08
  Quality loops that run themselvesovermanual bug triage

Frequently asked

## Questions, answered.

**What is Agent-Driven Development?**

Agent-Driven Development (ADD) is a way of building software where specialised AI agents drive every phase of the lifecycle — intake, spec, design, tech architecture, implementation, testing, UAT, deployment, retrospective — and humans review, decide, and steer. Glideship is the open-source reference implementation.

**Is Glideship a self-hosted Jira alternative?**

Yes — for the workflow layer. Glideship sits between IDE-side AI coding assistants and traditional PM tools. It is especially relevant to teams looking for an Atlassian Data Center alternative, since Atlassian stopped selling Jira Data Center to new customers in March 2026 and Data Center reaches end of life in March 2029.

**Does my code leave my machine when I use Glideship?**

The data plane is always local — Postgres, embeddings, BUDs, scanned repos, and the audit log all sit on your hardware. Only the LLM prompts leave your machine, and only when you choose a cloud inference mode.

**What does BUD mean in Glideship?**

BUD stands for Business Understanding Document. Every feature lives in one BUD containing spec, tech spec, test plan, acceptance criteria, and full history — all in markdown.

**What is Glideship's licence? Can I use it commercially?**

Glideship is licensed under the Apache License 2.0, which permits commercial use including embedding in proprietary products. Contributions require DCO sign-off via git commit -s.

## See it run on your codebase

[Get started on GitHub](https://github.com/mickyarun/glideship)

One command. Needs Docker and Node.js 18 or later. Your code and data stay on your machine. `npx glideship init`
