Agentile
Agile, with agency.
A deliberately light methodology for small teams who build software with AI agents. It keeps the agile spirit — short loops, working software, responding to change — and adapts it for a team where humans and agents work side by side.
The Manifesto
We are building software in an age where agents write much of the code. Agile taught us to favour short loops, working software, and responding to change. That spirit still holds — but the team has changed. Now humans and agents build together.
When we automated the building, review became the bottleneck. Then review could be automated too. What is left — the thing that is now hardest, and matters most — is describing what we actually want. As the building becomes automatic, what you ask for becomes the most important.
Be careful what you wish for — a vague wish, faithfully built, is still the wrong thing. And careful not to wish for too little: when building is so cheap, asking for less than you want is the costly mistake. Be precise, and be ambitious.
We have come to value
That is, while there is value in the items on the right, we value the items on the left more.
The principles behind it
-
1
Put care into the ask. Saying exactly what you want is the hardest, most valuable thing — and it decides everything downstream. Capture, shape, prioritise, and plan exist to get the ask right before anything is built: precise, and not too modest.
-
2
Small asks, one trunk. The unit of work is what you ask for; keep each ask small and specific, and integrate continuously to one mainline. Small asks make small diffs — reviewable by a person or an agent, and safe to land even when many work at once.
-
3
The spec is the unit of work; capture is free. Ideas enter as one-line stubs with zero ceremony. Nothing is built without a shaped spec.
-
4
Stages are yours to define. Agentile defines the loop and what the stages are; what each stage contains is yours to define. An implementation gives you a way to tailor or replace any stage — but that mechanism is the tool's, not the methodology's.
-
5
Context is infrastructure. Decisions, conventions, and domain rules live where every agent reads them, versioned with the code.
-
6
Reduce agency with determinism. Agents get judgement where it matters; anything you would repeat — build, test, scan, deploy — runs as a deterministic gate, never a request in prose.
-
7
A fresh reviewer beats a confident author. Agents find others' mistakes better than their own. Nothing merges unverified.
-
8
Accountability stays human. Agents draft, plan, build, and review; a named person approves the plan and signs off before anything ships.
-
9
The loop learns. Each cycle encodes its lessons, so the next one is cheaper.
First be agile, then agentic. Get the trunk, the gates, and the written spec right first — agents make a good loop fast, not a bad loop safe.
Before anything is built
The ask is the work
The hardest, most valuable act is saying exactly what you want. Agentile makes that the main event — most of the loop happens before anything is built. Capture, shape, prioritise, and plan aren't pre-work; they are the craft of turning a rough wish into a precise, worthy ask.
- Capture
- Never lose a wish — any idea lands in one line, instantly.
- Shape
- Sharpen it: who it's for, what "done" looks like, what's out of scope, the outcome that proves it.
- Prioritise
- Spend your care on the asks that matter most.
- Plan
- Turn the ask into an approach you can still steer — the last cheap moment to correct course.
Then build, verify, ship, and learn carry it out and check it against what you asked for. Each step is human, agent, or both.
The Loop
Eight stages, one trunk
Work flows around a single loop — a cycle is hours to a day, not a two-week sprint. Each cycle ends by encoding what it learned, so the next one is cheaper.
Every stage can be tailored to how your project actually works.
Light by design
Agentile defines the loop and the stages; what each stage contains is yours. Your Definition of Ready, your gates, your sign-off points — tailored per project by whatever tool you run it with.
Deterministic gates
Build, test, scan, and deploy are commands the agents run — never hopeful sentences in a prompt.
Trust, but verify
A fresh-context reviewer critiques every diff before it merges. Accountability stays human.
Install
Run Agentile in Claude Code
Agentile ships as a Claude Code plugin — thirteen /ag- commands,
three agents, and two enforcement hooks. Add the marketplace, install, then initialise your project.
claude plugin marketplace add agenta-consulting/agentile
claude plugin install agentile@agentile
Then, in any project you want to run the loop in:
/ag-init # scaffold docs/agentile/ (inbox + specs), .agentile/ config, docs/adr/
/ag-capture # drop a one-line idea, mid-build
/ag-shape # interview it into a ready spec
/ag-prioritise # order the ready queue
/ag-loop # work the queue: claim → build → verify → your sign-off → ship
Skills — ag-init, ag-capture, ag-inbox, ag-shape, ag-spec, ag-plan, ag-prioritise, ag-next, ag-wip, ag-abandon, ag-loop, ag-customise, ag-retro
Agents — planner, builder, reviewer, each in fresh context
Hooks — auto-format on edit, block "done" until tests pass
Yours — every stage takes a playbook: tailor gates, shaping questions, and sign-off points per project