S — VContact ↗
Back to overview

AI-native engineering

A maturity model for responsible agentic product development.

The goal is not novelty. The goal is better software delivery — stronger review, clearer context, measured quality, and team learning.

  1. Level 1

    01

    Individual prompting

    Engineers use AI tools independently for explanation, drafting, coding support, or debugging.

    What it looks like

    • Prompting varies by person.
    • Outputs stay mostly private until work is reviewed.
    • Value is real but hard to repeat.

    Risks

    • Uneven quality
    • Hidden assumptions
    • Over-trust in generated output

    Manager responsibility

    • Set basic expectations for review, privacy, and test discipline.
    • Coach engineers to show assumptions and reasoning.

    Metrics

    • Adoption signal
    • Review findings
    • Defect patterns

  2. Level 2

    02

    Shared prompts and reusable workflows

    Teams create reusable prompts and workflow patterns for common planning, review, QA, and documentation tasks.

    What it looks like

    • Prompt libraries and examples are shared.
    • Team members compare outputs and improve patterns.
    • AI support becomes part of normal delivery conversations.

    Risks

    • Stale prompts
    • False consistency
    • Prompts detached from repository reality

    Manager responsibility

    • Keep workflows tied to team standards.
    • Make shared examples reviewable and easy to update.

    Metrics

    • Workflow reuse
    • Ticket readiness
    • PR review cycle quality

  3. Level 3

    03

    Repo-specific agent instructions and skills

    Agents use repository-specific context, conventions, and quality gates before proposing code or workflow changes.

    What it looks like

    • Repositories include explicit agent guidance.
    • Agents read relevant files before editing.
    • Work aligns better with local architecture and tests.

    Risks

    • Outdated instructions
    • Context gaps
    • Automation that hides complexity

    Manager responsibility

    • Assign ownership for instruction maintenance.
    • Use code review to validate that guidance improves outcomes.

    Metrics

    • Instruction freshness
    • Rework rate
    • Build and lint pass rate

  4. Level 4

    04

    Jira/GitHub/CI-integrated agents

    Agents participate in delivery systems by reviewing tickets, drafting implementation plans, preparing pull requests, and responding to CI signals.

    What it looks like

    • Ticket and PR workflows include agent-assisted checks.
    • CI feedback is summarized into actionable remediation.
    • Engineers retain ownership of final decisions.

    Risks

    • Noisy automation
    • Permission creep
    • Overloaded review surfaces

    Manager responsibility

    • Design clear boundaries for agent actions.
    • Measure whether integrations reduce or increase delivery friction.

    Metrics

    • Cycle time
    • CI failure recovery
    • CTA and review completion rates

  5. Level 5

    05

    Measured agentic SDLC with evals, observability, and governance

    AI-assisted delivery is measured, governed, and continuously improved with quality signals, evals, and operational feedback.

    What it looks like

    • Workflows have measurable success criteria.
    • Teams evaluate agent output quality over time.
    • Governance protects quality, privacy, and maintainability.

    Risks

    • Metric gaming
    • Governance theater
    • Treating evals as a substitute for judgment

    Manager responsibility

    • Tie metrics to delivery quality and team learning.
    • Keep governance lightweight, real, and aligned with risk.

    Metrics

    • Defect escape rate
    • Review quality
    • Workflow eval results
    • Operational noise