FPT SoftwareFHNCertificationsReference architectures
Talk to NGT
AI & Agents · 04 of 6

AI-assisted delivery

AI speeds up delivery; engineers own the design, the review and the sign-off.

AI-Native SDLC with Human Gates
01

An agent team, with people at the gates

A ticket becomes a tested change. Engineers approve the plan and merge the result; quality gates run between.

AI-Native SDLC with Human Gates
ticketplanchange + testspassedmergedTeam toolsEngineers decideAgentsGit and the pipelineMicrosoft TeamsdiscussionJiraticketsFigmadesignsMCPtools for agentsLead agentplans, coordinatesBackend agentTest agentArchitecture agentGit repositorycode, prompts, runbooksReview agentevery pull requestQuality gatesGitLab CI, testsAWS CDKinfrastructure as codeEnvironmentsSLOs, alertsPlan reviewapprove, edit or rejectReview and mergean engineer signs off

Point at a component to light its links; select it for what it does here and the practice that makes it right.

Point at a component to see what it does; select it for the best practice.Open the architecture
02

Three levels of AI in engineering

FHN's maturity model, which NGT teams work to. Each level changes who approves what.

  1. 01Level 1 · AI-assisted
    • Human-led; AI is a sidekick
    • Optional tools for single tasks
    • Peer review as usual
    a baseline to measure againstmeasure before scaling
  2. 02Level 2 · AI-integrated
    • AI in IDEs and pipelines
    • AI proposes, a person decides
    • Sign-off before commit or deploy
    measured adoption and impactenforced policies and logs
  3. 03Level 3 · AI-orchestrated
    • Agents run scoped tasks: auto-PRs, self-healing
    • People set intent and guardrails
    • Policy-gated approval, full audit
    continuous risk and drift management
03

Lessons learned: do and don't

Nine rules NGT teams follow when they use AI tools.

Don't Do
Tool choiceA general chatbot for coding.The tool built for the job: Copilot, Cursor, Claude Code, Amazon Q or CodeVista.
LicencesPersonal accounts for customer work.Enterprise licences, data sharing off.
PromptsVague prompts.Clear, step-by-step instructions.
Task sizeA whole feature in one prompt.Small pieces, so no step is skipped.
ContextEvery session from nothing.Reusable prompt sets and rules.
Customer dataCustomer information in prompts.Public data, or masked data.
PrecisionCommand-line tasks left to the agent.Exact files named; commands run by people.
VerificationOutput trusted unchecked.Every output reviewed and fact-checked.
Know-howWhat worked is thrown away.Lessons kept as prompt sets and a handbook.
SourcesNGT Service Offering v3.2 (Jul 2026)AI-First DevOps Transformation: Services & Solutions, NGT (Oct 2026)Full SDLC with GenAI & Multi-Agent AI, NGT (Jun 2025)Platform Engineering v1.2, NGT (Sep 2026)FHN AI Service Introduction v0.9.3 (Mar 2026)Executive briefing: Look Forward 2027, FHN (Oct 2026)FPT Smart Cloud: AI Agents, Token Factory and AI Factory introductions (2025–2026)