Shefing AI asset · Software engineering

Shefing iSDLC^™

Agile reworked for the agent era.

A 7-stage lifecycle in which humans own intent, risk and accountability, while agents execute against versioned, machine-readable context.

1Goals
2Requirements
3Design
4Architecture
5Coding / CI / Testing
6IaC / CD
7Observability

AI shifts the bottleneck from writing code to steering and verifying it.

Intent made explicitEvery deliverable verifiableA tamper-evident record of who decided what

The lifecycle

Seven stages. Humans hold the gates.

Agents draft, build, test and watch production. At every stage that carries risk, a named person decides, and that decision is recorded.

Goals

Human sets intent

Requirements

Human sign-off

Design

Agent-led, human review

Architecture

Human sign-off

Coding / CI / Testing

Agents build, tiered gates

IaC / CD

Human gate on data & security

Observability

Agent-led, human accountable

agent-led human gate

Main features

Runnable, not just documented.

Seven mechanisms turn "humans in the loop" from a slogan into something a team can run and an auditor can check.

Manifesto · 12 principles

Agile's values, amended

Human accountability and managed context over individuals and interactions
Verifiable software and agent-ready context over working software
Explicit intent over collaboration
Architectural fluidity over responding to change

Inner loop

Plan → Build → Review

Humans author the plan: objective, Gherkin acceptance criteria, a DO-NOT list, context refs. Agents execute inside those boundaries. A cross-model review closes the loop: FAIL blocks, and every WARN needs a scribed human approval.

Executable workflow

15 skills, shipped as a Claude Code plugin

The methodology runs, it isn't just documented.

audit
goals
requirements
spec
wireframe
architecture
plan
build
tdd
review
qa
security
redteam
ship
observe

Accountability spine

An append-only, hash-chained trail

Records adoption, decisions, review verdicts and deploys, and flags commits that arrive with no justifying evidence. Agents may propose. Only humans attest.

Context substrate

A fresh agent onboards in 60 seconds

STATE.md · SPEC.md · AGENTS.md · QUIRKS.md · ADRs · TECH_DEBT.md · llms.txt, kept current by rule. A doctor command verifies presence and freshness.

Two entry doors

Greenfield, or audit first

Start clean, or onboard an existing codebase by paying down context debt first: SBOM, CVE baseline, secret scan, license and EOL audit, before any agent gets repo access.

Guardrails

Deliberation is never automated away

Nothing auto-writes a plan, auto-approves a PR or bypasses a red gate without an explicit human override plus an ADR. Tiered gating scales ceremony to risk.

Regulation

A core foundation for the EU AI Act

The Act asks software teams for human oversight, traceability, risk management and documentation. iSDLC produces all four by construction, as a by-product of how work gets done.

The Act requires

iSDLC provides

Article 14

Human oversight

Humans own intent, risk and every gate. Agents propose; only humans attest, and no red gate is bypassed without a recorded override.

Article 12

Record-keeping and traceability

The hash-chained, append-only accountability trail records who decided what, review verdicts and deploys, tamper-evident by design.

Article 9

Risk management

Tiered gating scales ceremony to risk, dedicated security and red-team stages, and drift detection on commits without justifying evidence.

Article 11

Technical documentation

Versioned, machine-readable context: specifications, architecture decision records and state, kept current by rule and verified for freshness.

Obligations depend on the risk class of your system. iSDLC gives you the evidence and the controls; qualifying a specific system remains a legal assessment.

Bring iSDLC to your team

We run it on our own delivery and stand it up inside yours.

Talk to us
← Back to our AI assets

Shefing iSDLC™