Skip to content

Latest commit

 

History

History
400 lines (342 loc) · 13.8 KB

File metadata and controls

400 lines (342 loc) · 13.8 KB

Agentic SDLC: Development Workflow

How a feature travels from client request to production in an AI-native pod


The Journey at a Glance

CLIENT
  │
  │  "We need X"
  ▼
┌─────────────────────────────────────────────────────────────┐
│  DISCOVERY & SCOPING                   [Product Owner]       │
│  Workshop recording → agent-drafted stories → PO refines    │
└──────────────────────────┬──────────────────────────────────┘
                           │  Approved spec
                           ▼
┌─────────────────────────────────────────────────────────────┐
│  DESIGN                                [Designer]            │
│  Prompt → 3–5 UI variants → pick + refine → design tokens   │
└──────────────────────────┬──────────────────────────────────┘
                           │  Figma + token handoff
                           ▼
┌─────────────────────────────────────────────────────────────┐
│  SPEC REVIEW                           [Tech Lead]           │
│  Architecture check · risk flagging · rules update          │
└──────────────────────────┬──────────────────────────────────┘
                           │  Green-lit spec
                           ▼
┌─────────────────────────────────────────────────────────────┐
│  PARALLEL AGENT EXECUTION              [Engineers]           │
│                                                              │
│   Agent A          Agent B          Agent C                  │
│   API layer    DB migration      Tests + docs                │
│      └──────────────┴────────────────┘                       │
│                      │                                       │
│              PRs opened per subtask                          │
└──────────────────────┬──────────────────────────────────────┘
                       │  PRs ready
                       ▼
┌─────────────────────────────────────────────────────────────┐
│  CODE REVIEW                           [Tech Lead + Eng]     │
│  Agent PR audit · architecture check · merge or revise      │
└──────────────────────┬──────────────────────────────────────┘
                       │  Merged to staging
                       ▼
┌─────────────────────────────────────────────────────────────┐
│  QUALITY ASSURANCE                     [QA]                  │
│  Agent test suite · exploratory testing · acceptance check  │
└──────────────────────┬──────────────────────────────────────┘
                       │  QA sign-off
                       ▼
┌─────────────────────────────────────────────────────────────┐
│  RELEASE                               [DevOps / Platform]   │
│  Canary → 5% traffic · metrics watch · full rollout         │
└──────────────────────┬──────────────────────────────────────┘
                       │
                       ▼
                  PRODUCTION
                       │
                       │  Demo + sign-off
                       ▼
                    CLIENT

Stage 1 — Discovery & Scoping

Owner: Product Owner
Inputs: Client brief, workshop recordings, existing backlog
Outputs: Refined user stories, acceptance criteria, scope boundary

Client call / workshop
        │
        ▼
  Otter / Fireflies
  (recording + transcript)
        │
        ▼
  Agent drafts stories
  from transcript
        │
        ▼
  PO reviews + fixes
  (2 of 8 stories need edits)
        │
   ┌────┴────┐
   │         │
In scope   Out of scope
(backlog)  (logged as future / change order)

What agents do: Transcribe, structure raw notes into stories, score priority, flag ambiguity.
What only humans do: Read client mood, make scope trade-off calls, protect margin.

Definition of done: Every story has a title, user value statement, acceptance criteria, and a clear "not in scope" boundary. Coding agents can act on it without asking follow-up questions.


Stage 2 — Design

Owner: Designer
Inputs: Approved stories, design system tokens, brand guidelines
Outputs: Refined Figma screens, updated design tokens, prototype if needed

Story + brand context
        │
        ▼
  Prompt to Claude Design / v0
        │
        ▼
  3–5 UI variants generated
        │
        ▼
  Designer picks + refines
        │
   ┌────┴──────────────┐
   │                   │
Figma handoff     Token update
(screens)         (colors, type,
                   spacing in code)

Checkpoint: Tech Lead reviews for implementability before engineering starts. Catches layout patterns that would be expensive to build before a line of code is written.


Stage 3 — Spec Review

Owner: Tech Lead
Inputs: Refined stories, design handoff
Outputs: Go/no-go on spec, architecture notes, updated CLAUDE.md

Story + Figma
      │
      ▼
  Tech Lead reads spec
      │
  ┌───┴───────────────┐
  │                   │
Good to go       Needs work
  │                   │
  ▼                   ▼
Update           Bounce back
CLAUDE.md        to PO or Designer
with any         with specific gaps
new rules

This is where quality is won or lost. A vague spec produces vague agent output. The Tech Lead's job here is to make the spec unambiguous before any code runs.


Stage 4 — Parallel Agent Execution

Owner: Engineer(s)
Inputs: Green-lit spec, CLAUDE.md, design tokens
Outputs: Feature branches, opened PRs per subtask

Engineer reads spec
        │
        ▼
  Splits into independent subtasks
        │
   ┌────┼────┐
   │    │    │
   ▼    ▼    ▼
Agent  Agent  Agent
  A      B      C
  │      │      │
  ▼      ▼      ▼
API   DB +    Tests
layer  migration  + docs
  │      │      │
  └────┬─┘      │
       │        │
       ▼        ▼
     PR A+B    PR C

Engineer's role: Write the specs, spawn the agents, monitor for drift, step in early when an agent goes off course, review every PR as if a junior wrote it.

Cost gate: Engineer checks token burn after the first agent cycle. A runaway loop can cost $400+ before anyone notices.


Stage 5 — Code Review

Owner: Tech Lead (final gate) + Engineer (peer)
Inputs: Agent-generated PRs
Outputs: Merged code, updated rules for repeat mistakes

Agent PR opened
       │
       ▼
  Automated checks
  (lint, types, tests)
       │
   ┌───┴────────┐
   │            │
Passes       Fails
   │            │
   ▼            ▼
Engineer     Agent or
peer review  engineer
             fixes
       │
       ▼
  Tech Lead review
  (architecture, security,
   subtle logic)
       │
   ┌───┴────────┐
   │            │
Approved     Changes
   │          requested
   ▼
 Merge
   │
   ▼
Update CLAUDE.md
if a new rule would
prevent this class
of mistake

What to watch for in agent PRs: Plausible-looking but subtly wrong logic, slow or unsafe queries, missing auth checks, over-engineered solutions.


Stage 6 — Quality Assurance

Owner: QA
Inputs: Feature on staging, acceptance criteria from Stage 1
Outputs: QA sign-off or bug report

Feature on staging
        │
   ┌────┴─────────────────┐
   │                      │
Agent test suite      Exploratory
(Playwright, 140+     testing by human
 tests overnight)     (edge cases, UX
        │              feel, business
   ┌────┴────┐          logic sense)
   │         │               │
Auto-fix   Escalate          │
 (healer    to QA     ┌──────┘
  agent)    or eng    │
                 Acceptance criteria
                 check against spec
                      │
                 ┌────┴────┐
                 │         │
              Pass        Fail
                 │         │
              Sign-off   Bug report
                         → back to
                           Stage 4

For AI-powered features: QA also runs LLM-as-judge evaluation — Braintrust or LangSmith scoring agent output against rubrics, not just checking UI clicks.


Stage 7 — Release

Owner: DevOps / Platform
Inputs: QA sign-off, merged code on main
Outputs: Feature live in production

Merge to main
      │
      ▼
  CI/CD pipeline
  (build, scan, smoke test)
      │
      ▼
  Canary deploy
  (5% of traffic)
      │
  ┌───┴───────────────┐
  │                   │
Metrics OK         Anomaly
(latency,           detected
 error rate,           │
 cost)             Auto-rollback
  │                + alert
  ▼
Full rollout
(100% traffic)
      │
      ▼
Observability live
(LangSmith traces,
 Helicone cost,
 Grafana metrics)

Non-deterministic release strategy: Feature flags let DevOps roll out AI-powered features to a slice of users first, measure quality, and expand — or kill — without a full redeploy.


Stage 8 — Client Demo & Sign-off

Owner: Product Owner
Inputs: Feature live in production
Outputs: Client approval, invoice milestone hit, or change order

Feature live
     │
     ▼
  PO prepares demo
  (agent drafts talking points,
   PO edits for client tone)
     │
     ▼
  Demo call
     │
  ┌──┴──────────┐
  │             │
Approved    Feedback
  │             │
  ▼          Triage:
Invoice         │
milestone    ┌──┴──────────┐
hit          │             │
          In scope     Change order
          (bug fix)    (new scope,
                        new SOW)

Timeline: What a Typical Feature Looks Like

Day 1   ████  Discovery & Scoping (PO + agents + client)
Day 2   ██    Design (Designer + AI tools)
Day 2   █     Spec Review (Tech Lead)
Day 3–4 ████  Parallel Agent Execution (Engineers + agents)
Day 4–5 ██    Code Review (Tech Lead + Engineers)
Day 5   ██    QA (automated overnight + human exploratory)
Day 6   █     Release (canary → full rollout)
Day 6   █     Demo + sign-off (PO + client)

A feature that took 3–4 weeks in the old model ships in 5–6 days. The constraint is no longer writing code — it is judgment: the PO reading the client, the Tech Lead spotting the subtle bug, the QA engineer finding the edge case no agent imagined.


Who Does What — Summary

Stage Human owner Agent contribution Human judgment required
Discovery & Scoping Product Owner Draft stories from transcripts Scope calls, client trust, margin
Design Designer Generate UI variants Brand, accessibility, consent UX
Spec Review Tech Lead Architecture, risk, ambiguity
Development Engineer Write code, tests, docs Spec quality, drift detection
Code Review Tech Lead + Eng Automated checks Subtle logic, security
QA QA Test generation, auto-healing Exploratory, business sense
Release DevOps Anomaly detection, rollback Risk tolerance, canary thresholds
Sign-off Product Owner Demo notes Client relationship, scope judgment

Feedback Loops That Keep Quality High

Repeat agent mistake
         │
         ▼
    New rule in CLAUDE.md
         │
         ▼
    All future agent runs
    read the updated rule
         │
         ▼
    Whole class of bug
    prevented, not just
    one instance fixed

This is the compounding advantage of the agentic model: every caught mistake that becomes a rule makes every future feature better. The CLAUDE.md file is the pod's institutional memory.


The workflow is not waterfall — stages overlap, agents run in parallel, and fast features can skip from spec to staging in a single day. The diagram shows the logical sequence, not a rigid calendar.