AI coding agents can lose context, drift beyond scope, and mistake passing tests for completed work. Agentic Development gives them a repeatable workflow: clear goals, bounded changes, verifiable evidence, and explicit handoffs.
Read the getting started guide, then give your agent a concrete outcome:
Use Agentic Development to deliver this change: [desired behavior]. Clarify scope and acceptance criteria before editing. Verify the result, report remaining risks, and ask before actions that need my approval.
Seven reusable Skills guide the work: six phases and one coordinator.
Goal / need → Research → Plan → Implement → Test → Judge → Scoped acceptance
↑ │
└── Iterate ←─────┘
if correction is needed
| Phase | Question to resolve | Expected output |
|---|---|---|
| Research | What must we understand before deciding? | Cited findings, risks, unknowns, and constraints. |
| Plan | What will we deliver, and how will we recognize success? | Scope, acceptance criteria, dependencies, tasks, and review policy. |
| Implement | How do we execute the authorized task without scope drift? | A focused diff, preserved existing work, and initial verification evidence. |
| Test | What does this exact candidate actually demonstrate? | Criterion-linked checks, observed results, and coverage/environment limits. |
| Judge | Is the evidence sufficient to accept this scope? | A verdict on a frozen candidate, evidence-linked findings, and the agreed evaluation mode. |
| Iterate | What is the smallest correction that addresses the failures? | Bounded correction, followed by renewed verification and judgment. |
cycle coordinates these transitions: it selects the needed phase, verifies prerequisites, and prepares resumable handoffs.
Keep requirements, decisions, and restart context explicit; make small, reviewable changes; separate test results from acceptance and report missing evidence. Use the whole cycle or the phase your task needs. These are agent instructions, not an enforcement engine or a guarantee of correct software.
Install: goose plugin install https://github.com/bioinfornatics/agentic-development.git.
Check: goose skills list — only this plugin’s seven Skills, no evaluation fixtures.
Update: goose plugin update agentic-development. See Goose Plugins and Agent Skills.
The Skills are Markdown instructions: use an agent that can load them and your project's own development tools. No package installation or build is required for the Skills.
| Tool | Why use it? | How it fits into the workflow |
|---|---|---|
| Git | Preserve reviewable source history. | Inspect the diff, then commit verified changes when authorized. |
GitHub CLI (gh), optional |
Manage GitHub reviews from the terminal. | Verify the account and repository; create or inspect PRs and check CI results. |
Beads (bd), required when the project uses it |
Recover task scope, ownership, dependencies, and progress across sessions. | Read and claim the task before editing; record results and the next action in the existing tracker. |
| Entire, optional | Capture session context and connect it to checkpoints. | Verify support for your agent and worktree; link only observed session/checkpoint IDs to the task. |
Handoff: user prompt → task with scope and acceptance criteria → execution → evidence, status, and next action recorded in Beads → restart prompt with task ID. The next session rereads the task: Beads is task truth, not the prompt or a notebook. Beads adoption is conditional; once chosen, missing access or a failed claim blocks implementation, not permission to continue untracked. Without Beads, agree another tracking approach. Entire captures context, not task truth or acceptance; no tool grants permission to commit, push, or close tasks. See the integration guide for setup checks and boundaries.
- Development cycle: how the phases fit together.
- Integrations: Beads, Git, Entire, and optional Goose observation hooks.
- Contributing: improve the method on develop.
The main branch contains the generated installation package; develop contains its source and documentation. Both use this same README. Skills need no build; optional observation hooks require Python 3.10+ and use only the standard library.