EUWardenAI SDLC
Process architecture Process tools Contact
A multi-domain software delivery process with AI agents

AI agents build.
Gates prove.
Humans decide.

AI SDLC is a complete software development lifecycle in which specialized AI agents carry work through nine stages — and every transition between stages requires evidence, not declarations. One process for many domains, a full audit trail in Jira and GitHub, and a human owning every strategic decision.

9lifecycle stages — from a stated need to post-release lessons
7quality gates — each with one accountable role and machine-verifiable evidence
1unified process for initiatives, domain work packages, and small changes
100%of transitions backed by evidence pinned to an exact code revision
01Design principle

Autonomy without control is debt. Control without autonomy is stagnation.

AI agents writing code without a process produce fast — and unpredictably. AI SDLC separates two powers: merit (who creates the stage's content) and flow (who moves work forward). No role holds both at once.

Principle 1

One lifecycle

A business initiative, a domain work package, and a small fix all pass through the same stages. The work type changes the stage's output — never the process itself. Zero exceptions, zero side doors to "Done".

Principle 2

Domain and process are orthogonal

The domain answers the question where work lands and who decides. The process answers the question what happens next. No domain gets a lifecycle of its own.

Principle 3

Transitions carry no decisions

A status change is pure mechanics: it fires only once gate evidence exists. The transition script verifies evidence — it never creates it. The merit decision is made earlier, by the right role.

This is not just our observation. The 2025 DORA report (Google Cloud, ~5,000 respondents) finds AI adoption now correlates positively with delivery throughput — but still negatively with delivery stability. Code arrives faster and changes break more often. DORA calls AI an amplifier: it magnifies an organization’s strengths and its weaknesses alike. DORA’s own conclusions point to small batches, automated testing and fast feedback loops as the counterweight. The separation of powers described below is our answer to that diagnosis — it is not a DORA recommendation. Full figures and sources are on the Results page.

02Lifecycle

Nine stages. One role per stage. One exit gate.

Click a stage to see its contract: who is accountable, what gets produced, and what evidence opens the transition forward.

03Gates and evidence

Evidence, not declarations

Every stage ends with a single line of machine-format evidence, published in Jira by the accountable role — and only by that role. The evidence is pinned to an exact code revision (a 40-character commit SHA): the verdict covers what will actually ship, not "roughly this version".

The system runs fail-closed: evidence that is missing, ambiguous, duplicated, or pinned to a different revision means the transition is refused with a machine-readable refusal packet. There is no "pushing" a status through: the transition script checks the evidence live in Jira, every single time.

When the code changes after a verdict is issued, all evidence expires together and the verification stage runs again on the new revision. This closes the classic gap of "we tested it, then someone pushed one more commit".

StageAccountable roleGate evidenceTransition fired by
AnalysisAnalyst AgentANALYSIS GATE: PASSDirector
ArchitectureArchitect AgentARCHITECTURE GATE: PASSDirector
Implementation planningPlanner Agent (solution architect)IMPLEMENTATION PLANNING GATE: PASSDirector
DevelopmentRouted developer agentDEVELOPMENT GATE: PASSDirector
Verification (QA)QA AgentVERIFICATION GATE: PASSQA Agent
Release / CloseRelease AgentRELEASE GATE: PASSRelease Agent
LearningProcess OwnerLEARNING GATE: PASSProcess Owner

Conditional verdicts: architecture and security

When a change is architectural in nature or touches a security surface (authentication, cross-tenant data, secrets, new dependencies…), the Verification stage requires additional specialist verdicts — all pinned to the same revision. QA refuses PASS until the full set exists. When in doubt, the classification tilts toward "required". This is audit-ready evidence for SOC 2 / ISO 27001.

VERIFICATION GATE QA Agent ARCHITECTURE VERIFICATION Architect Agent · on architecture change SECURITY VERIFICATION Security Agent · when security is touched Verification artifact all verdicts + CI + review HEAD SHA 4f9c2e7…
All required verdicts bind to one exact revision. A code change invalidates the whole set.
04Rework paths

Defects go back to the source, not to a patch

When a gate refuses, work does not "fall to the developer by default". It returns to the earliest stage whose accepted contract turned out to be wrong: a flawed requirement — to Analysis, a flawed design — to Architecture, a flawed implementation — to Development. A later role never quietly patches someone else's contract.

Analysis Architecture Planning Development Verification Release flawed design → Architecture flawed requirement → Analysis flawed implementation → Development plan does not hold → Planning revision drift → re-Verification flawed requirement flawed design
Seven named return paths — each with a reason recorded in Jira. The stage that detects a defect sends it back: Architecture can return a requirement to Analysis, Planning a design to Architecture — without waiting for Verification. Returned work goes through Development and Verification again, on a new revision. The "Blocked" state is separate: it remembers the return stage and deterministically resumes exactly there.
05Roles and separation of powers

Every role is a specialized skill with its own contract

An agent role is not a "general prompt" — it is a persistent, versioned skill with an input/output contract, permissions limited to the task, and a ban on stepping outside its assigned scope. Roles fall into two disjoint classes.

Flow authority — orchestration, never content

Operator (human)

Owner of all domains. Decides scope, priority, risk, and result acceptance. The sole source of strategic decisions.

Initiative Director

Temporary, for the span of one initiative. Coordinates sequencing and cross-domain dependencies; performs no merit work.

Domain Director

Permanent, one per domain. Runs the domain queue, fires transitions based on evidence, delegates all content to merit roles.

Domain Steward

Guardian of coherence: assesses domain health against its strategy and boundaries. Recommends corrections — does not run the queue and does not change scope.

Merit authority — stage content, never status

Analyst Agent

Turns a need into accepted scope: requirements, testable acceptance criteria, estimation, risks.

Architect Agent

Architecture decisions (ADRs), threat model, system boundaries — or an explicit, justified "N/A". Architecture is always visited.

Planner Agent

The implementation package: work units, target paths, specialist selection per unit, concurrency, verification commands.

Developer Agents

Specializations: backend, frontend, browser testing, DevOps. Deterministic routing by file paths; one is accountable for the merged result.

QA Agent

The sole author of the verification verdict and the only role firing the transition out of QA. Consumes specialist verdicts; never issues them itself.

Security Agent

SAST/SCA scans, threat model validation, the security verdict. Evidence for QA — not authority over status.

Release Agent

Mechanical execution: merge only through the deterministic merge gate, post-deployment runtime verification, cleanup.

Process Owner

Runs the Learning stage: a retrospective for every delivery, lessons turned into process improvements with an assigned owner.

MERIT ROLE does the stage work, judges its own contract never changes status TICKET <GATE>: PASS <sha> one authoritative line durable, signed record TRANSITION GATE checks form and binding to the revision, live zero merit judgment NEXT STATUS exact-name transition from the live menu state changes once authors reads fires any check fails REFUSAL PACKET — no transition machine-readable, non-zero exit code the default outcome is refusal, not a pass
Three parties at every transition: the author of the verdict, the verifier of its form, and the firer. The Director fires through Development, the named role afterwards — but the gate trusts only what it reads from the ticket itself, never the caller's claim. Nobody grades their own work at transition time.

The hard boundary: a Director never creates merit content — no analysis, architecture, code, or verdict. A merit role never changes status in Jira. As a result, no single agent can both "do and approve" — the same principle that separation of duties in financial auditing is built on.

06Multi-domain model

One initiative, many domains, zero chaos

A business initiative goes through its own cross-domain analysis and architecture, then decomposes into Domain Work Packages (DWPs) — durable delivery units, each belonging to exactly one domain and passing through the full lifecycle at its own pace. Purely domain work (debt, maintenance, small changes) enters the domain queue directly — with no artificial initiative. A domain is an area of business ownership, not a technical layer: the interface, API and database of one change all belong to the same domain.

INITIATIVE business goal · Initiative Director cross-domain analysis + architecture → DWP decomposition DOMAIN · Sales DOMAIN · Integrations DOMAIN · Infrastructure DOMAIN · AI SDLC Domain Director · queue Domain Director · queue Domain Director · queue Domain Director · queue Steward — domain coherence Steward — domain coherence Steward — domain coherence Steward — domain coherence DWP · lead scoring DWP · SIEM export DWP · deploy pipeline DWP · new QA gate DWP · quote pipeline DWP · calendar sync domain work (no initiative) full lifecycle in domain full lifecycle in domain full lifecycle in domain full lifecycle in domain cross-domain dependency = explicit contract, not a shared ticket
Example: a CRM system. The domain answers: where and who. The process answers: what next. Initiative and Domain Directors are peer roles optimizing different dimensions — not a hierarchy. A discovered impact on another domain escalates to the owner of its contract instead of quietly expanding the package. Real systems have more than four domains; the delivery process itself is one too — with its own Director, queue and full change lifecycle.
decomposition and work flow contracts, coherence, dependencies
07State systems integration

Jira as the ledger of state, GitHub as the ledger of code

The process is tool-independent — Jira materializes state, it does not define it. Every stage, gate evidence line, verdict, and operator decision lands as a durable, role-signed record on the issue. Below — a real run of one issue, step by step.

PRJ-1042 · Domain Work Package
Audit log export to the customer's SIEM
Intake
arch-verify-required sec-verify-required ai-wip:analyst
Transitions

A script, not a click

Statuses are changed only by the transition script, which verifies gate evidence live in Jira. Human and agent see the same state; nobody "drags cards" by hand.

Ownership

Claims, not queues

Who is currently working is recorded by a claim label — ownership metadata, not an extra status. Zero fictional "waiting for handoff" states.

GitHub

Merge only through the gate

Code ships exclusively through pull requests from isolated branches. The merge is executed by a deterministic merge gate once the evidence set is complete — no agent merges "by hand".

08Human in the loop

Closure never happens without human acceptance

Agents do the work and prove its quality — but accepting the result belongs to the Operator. An issue does not reach "Done" until a recorded human acceptance exists at the Learning stage. Scope, priority, risk, and domain boundaries are decisions reserved for humans; an agent that discovers them escalates — it never decides alone.

The Learning stage closes the loop: every delivery ends with a retrospective, and a detected process defect gets fixed or becomes a separate task with an owner. A process that measures and improves itself — that is the difference between "we use AI" and "we run a system that delivers better every week".

Delivery Learning retrospective → improvements lessons better process OPERATOR result acceptance
Without a recorded Operator acceptance, the Learning gate does not let work through to "Done".
What you get

Two perspectives, one system

For the manager

  • Predictable status without asking the team — the state of every piece of work is visible in Jira; "where is it" stops being a meeting.
  • SOC 2 / ISO 27001-grade auditability — every verdict, author, and revision recorded durably; audit evidence is produced as a by-product of the work.
  • Parallel scaling without chaos — many domains and many agents at once, because ownership and boundaries are explicit and collisions are impossible by definition.
  • The end of "AI wrote it, nobody checked" — no change reaches main without independent verification and the full set of gates.
  • A process that learns — a retrospective after every delivery turns incidents into improvements with an owner and a deadline.

For the architect

  • Architecture is always visited — no change skips the architecture stage; a deliberate "N/A" is a recorded architect decision too.
  • Clean stage contracts — one role, one output, one gate; stage inputs and outputs are explicit and testable.
  • Decisions in ADRs, verified on the diff — recording the decision is only the start: for significant changes the architect confirms the implementation matches the decision on an exact revision.
  • Domain boundaries protected by escalation — impact on someone else's domain is an explicit contract and an escalation, not silent scope creep.
  • Repair at the source — a requirements defect returns to analysis, a design defect to architecture; the system never entrenches patches on bad foundations.
Next step

See the process on a live system — from request to deployment in a single session.

This process is not a slide-deck concept. It runs in production: it drives real deliveries — including a security product built to audit standards — with real gates and a real evidence history. We would be glad to walk your team through the full cycle on an example of your choice. We run adoption itself as a process too: three phases with hard exit conditions, where "no-go" is a legitimate outcome that ends with a report, not a sanction — and retrospectives are about the process, never about people.

Role names, domain names, and issue identifiers in this presentation are examples. The gate evidence format and transition mechanics reflect a working system.