AGENTIX · ENGINEERING UPGRADEPART 01 / SDLC FRAMEWORK
Engineering Upgrade › SDLC Framework
§ Part 01 · how the work gets built

The agentic SDLC,
gated by people.

Twelve specialist agents carry a feature from a plain-English requirement to a merged PR, across six phases. The agents draft and recommend, a human approves every phase, and every call is costed. It lives in the repo, Claude Code is the UI, there's no new dashboard to learn.

FIG.01, the agentic SDLC · a human approval gate at every phase
P1 ⛓ Refine shape the requirement
P2 ⛓ Plan architecture + approach
P3 ⛓ Stories acceptance criteria
P4 ⛓ Implement build to the plan
P5 ⛓ PR reviewed + merged
Orange gates are the product, an agent drafts, recommends a path, but a person edits status: draft → approved before the next command runs. Commands abort if the upstream phase is not approved.
FIG.02, the command surface · one slash command per phase
CommandInvokesProduces
/sdlc-refine Orequirements-architect requirement.md
/sdlc-plan Osolution-architect plan.md
/sdlc-designoptional Oui-ux-designer design.md
/sdlc-stories Sstory-writer + estimator stories/NN
/sdlc-implement Oimplementer code + tests
/sdlc-pr Sreviewer + 3 stack reviewers pr.md + PR
/sync Sstatus-sync state rollup
/cost cost-logger hook feature cost report
Seven core commands run the lifecycle; /code-review, /portfolio and the roadmap commands sit alongside. The whole surface is invoked from inside Claude Code, in the repo.
FIG.03, the agent workforce · 12 specialists, routed by model tier
Opusheavy synthesis, design trade-offs, code correctness & safety calls Sonnetstructured transformation & checklist review — ~3× cheaper per call
Opus requirements-architect

Senior BA. Asks 4–10 clarifying questions, never guesses.

→ requirement.md
Opus solution-architect

Maps the requirement to modules, migrations and a file checklist.

→ plan.md
Opus ui-ux-designer

Designs screens with healthcare guardrails. Read-only, writes no code.

→ design.md
Sonnet story-writer

Splits the plan into vertical-slice stories with Given/When/Then ACs.

→ stories/NN
Opus implementer

Writes production code + co-located tests, one story per call.

→ code + tests
Sonnet reviewer

Orchestrates three stack reviewers, blocks the PR on any blocker.

→ pr.md + PR
Sonnet backend-reviewer

CQRS discipline, tenant isolation, validation, PHI rules.

→ verdict
Sonnet frontend-reviewer

React/TS patterns, state boundaries, accessibility, tests.

→ verdict
Sonnet aws-infra-reviewer

Least-privilege IAM, secrets, blast-radius, observability.

→ verdict
Opus product-auditor

Retrospective friction + AI-opportunity audit of shipped work.

→ audit.md
Sonnet status-sync

Propagates state bottom-up: story → feature → epic.

→ state
Sonnet dashboard-generator

Renders the live portfolio overview from current state.

→ overview.html
Routing is the cost story: synthesis runs on Opus, review and transformation run on Sonnet. A typical feature spends ~$0.86 of every $14 on its two Sonnet phases.
§ Serves six roles, not just developers
PM / BA

Owns the front of the funnel, refine to hand-off.

Architect

Maps requirement to architecture, owns the trade-offs.

Developer

Implements one story at a time, opens the PR.

QA

Validates acceptance criteria, signs off pre-release.

DevOps

Pipelines, secrets, infra and deploy gates.

Leadership

Velocity, spend and decisions, visible at a glance.

§ What changes for each role
RoleTodayWith the framework
PM / BA Hand-writes specs, scope is a guess Refines with an agent, scope is explicit and approved
Architect Trade-offs live in someone's head Captured in plan.md and cited ADRs
Developer Re-explains the domain every prompt One story, full context, co-located tests
QA Acceptance criteria checked by eye Encoded as Given/When/Then, verified
DevOps Infra reviewed manually, if at all aws-infra-reviewer gates every PR
Leadership No cost data, decisions on gut feel Cost per feature, captured from day one
FIG.04, a day in the life · PM runs /sdlc-refine, an agent does the rest
claude-code — care-platform
~/care-platform $ /sdlc-refine discharge-summary-redesign
→ requirements-architect (Opus) loading ADRs + module READMEs…
The current discharge screen has 47 fields. Only 8 are used in 90% of cases. Four clarifying questions:
1. Should unused fields be removed or hidden behind "advanced"?
you Hidden. Keep them for audit, collapse by default.
Understood. Drafting requirement.md with 6 dependency-ordered stories…
✓ requirement.md written · status: draft → awaiting PM approval
Hand-off complete. PM did zero copy-paste.
Non-negotiables, PHI protection, tenant isolation, CQRS discipline, module boundaries, are enforced by reviewer agents at the framework level, not buried in a wiki.
FIG.05, tokens & cost · measured from day one sample feature
feature: care-plan-v2Care Plan v2 — tenant portal · 6 stories, 1 PR merged
TOTAL$14.01
PhaseAgentCallsInputOutputCost
Refine Orequirements-architect 3 42k 8k $1.23
Plan Osolution-architect 2 68k 12k $1.92
Design Oui-ux-designer 1 35k 9k $1.06
Stories Sstory-writer + estimator 2 52k 7k $0.27
Implement Oimplementer ×6 stories 6 312k 48k $8.94
PR Sreviewer + 3 stack reviewers 4 128k 14k $0.59
6 stories · 1 PR 18637k98k$14.01
One logging hook records every agent call, per feature, per phase, per agent, per model, aggregated into feature.md on sync, with an anomaly alert if a feature exceeds 2× the median for similar-sized work. The full per-feature ledger feeds ROI reporting →
§ Four principles the framework holds to
01

Gates are the product

The agent drafts and recommends a path. A person edits status to approved before anything proceeds.

02

Commands abort upstream

No phase can skip its gate, /sdlc-plan refuses to run until the requirement is approved.

03

Colour is meaning

Draft, code, review, done, every status is legible at a glance, never decorative.

04

It lives in the repo

Claude Code is the UI. No new dashboard, no portal, no tool for anyone to learn.

← Back Engineering Upgrade overview Next → Intelligence (Jaadu)