Do it yourself
This week.
The first layer is free: the one page, a pre-filled table for your kind of team, the record template, the context-file template, and ten tests for your agent.
If it's running by Friday, you didn't need us.
AI-SDLC
You have the licenses.Now you need the method.
Your team has the agents. It doesn't yet have the way of working that turns them into faster, safer delivery.
Works with the tools you already have · Nothing to buy · Nothing hosted
Which one is you?
“Coding got faster. Delivery didn't.”
Agents everywhere, pull requests piling up, seniors are the queue, releases no quicker.
Something is missing. It isn't a tool.
“We bought the licenses. The vendor is training the team. We'll adapt our own process.”
Twenty years of shipping software. You've absorbed every shift before this one.
You're right. The question is what you're adapting to.
The evidence
The teams in these numbers had licenses, training and experienced leaders.
92%
were confident their AI-written code was production-ready. 81% then saw more production issues from it. Only 12% have anyone governing it.¹
32.7%
is the merge rate for AI-assisted pull requests, against about 84% for human-written ones. They wait five times longer for review.²
42%
of committed code is already AI-written or assisted, heading for 65% by 2027. Most teams can't say which changes.³
That is not the variable. The variable is whether anyone wrote it down.
Any agent
Works with the agents you already use.
- GitHub Copilot
- Claude Code
- OpenAI Codex
- Cursor
- Gemini CLI
- Windsurf
- Kiro
- GitLab Duo
- Devin
And whichever agent your team picks next
The method sits around your tools, not inside them. It runs on GitHub, GitLab, Azure DevOps or Bitbucket, with any AI model. Change agents next quarter and nothing in the method changes.
The method
The whole method fits on one page.
People specify, verify and operate. Agents build and test.
01
Five steps
Specify → Execute → Verify → Decide → Learn
02
Four levels
SHIP · SHOW · ASK · STOP, one per kind of work
03
One table
Your kinds of work and their levels, enforced in your branch rules
04
Three hats
Own the table · judge the evidence · run the agents. No new titles.
05
A brief with non-goals
Nothing starts without what's out of scope
06
A six-field record
What, kind, agent, checks, who decided, when
07
One context file
Under 100 lines. Every tool reads it.
08
One weekly meeting
Failures become fixes. One ships every week.
09
Three numbers
Agent share of work · cost per change · guardrails held
SHOW, the one people ask about
The agent merges when the checks pass. A named person decides within hours, before any customer sees it. The merge is not the release.
Four things a license can't give you.
CLEAR RULES FOR EACH TYPE OF WORK
For example: agents can add tests on their own, database changes need a person's approval first, and payments stay with people.
TRUST THAT IS EARNED
Agents get more freedom only after they prove themselves on your real work.
A NAME ON EVERY CHANGE
Someone always signs off, and you can see who, when and why in minutes.
HONEST NUMBERS
Faster only counts if quality, reliability and security didn't get worse.
Who this is for.
FOR
- Engineering teams of any size. We start with one team that has a real backlog and a pipeline, then scale.
- Leaders who have the licenses and want the return
- Global capability centres measured on AI productivity
- Regulated teams whose auditors are starting to ask
NOT FOR
- Teams without tests or CI. Build the foundation first.
- Anyone wanting agents in production with no human decision
- Anyone wanting a guaranteed productivity multiple
- Side projects and throwaway prototypes
Getting started
Three ways to start, from an afternoon to two weeks.
Ship Day
Spend a day with us.
We come to your repository, do the first three practices with your team, wire the table into your branch rules, and see the first merge-on-green with a named person deciding. Same day.
Diagnostic
Best place to start
Two weeks to a verdict.
Your baseline, from your systems. Agents on 8–12 live items in shadow, judged by your engineers. A written verdict: go · fix first · not yet.
Pricing on request. No outcome-linked fees: an advisor paid on the number can't be trusted to write “not yet”.
Self-check
Seven questions. Two minutes.
1Do you know how much of last sprint's code was written by AI agents?
2Have you written down which work agents may do on their own, which needs approval, and which they must never touch?
3Are those rules enforced by your tools (branch rules, code owners, permissions), not just written in a document?
4For any recent change, can you see in a few minutes who asked for it, what the agent did, which checks ran and who approved it?
5Have you tested your agents on your own codebase, including whether a hidden instruction in a ticket can trick them?
6Are payments, logins, customer data and your build pipeline off-limits to agents, with no access granted?
7Could your team keep delivering if you switched every agent off tomorrow?
The whitepaper
Six Weeks to an AI-SDLC
Twelve pages. The evidence, the model, the six weeks, and what it won't do.
Who we are.
AI for Growth. Independent, senior-led, based in Bengaluru, working with teams in the USA, the UK, India and the Gulf.
The method is called DIRIGENCE. The name matters less than the practice.
Built on ideas we credit: Ship / Show / Ask · DORA · AWS AI-DLC · maker-checker · permit-to-work