Skip to content

DORA 2025 AI Capabilities — Fawkes Alignment Guide

Source: DORA 2025 State of AI-Assisted Software Development
Category: Reference | AI Practices
Audience: Platform engineers, AI coding agents, team leads

Overview

The DORA 2025 State of AI-Assisted Software Development report established the DORA AI Capabilities Model: seven technical and cultural foundations that determine whether AI accelerates or destabilises software delivery. Approximately 90% of technology professionals now use AI in their work, yet ~30% express little or no trust in AI-generated code. Speed without stability accelerates chaos, not value.

Key finding: AI is an amplifier of existing practices, not a shortcut. High-performing teams get measurably better with AI. Teams with weak foundations see their problems worsen at greater speed.

The Seven Foundations — Fawkes Status

1. Clear AI Stance and Policy

Ambiguity breeds risk. Teams need defined, communicated policies on when and how to leverage AI safely.

Fawkes implementation: - AGENTS.md — universal rules for all agents (Copilot, Claude, VS Code agent mode) - .github/copilot-instructions.md — Copilot-specific standards - .github/agents/ — specialist agent profiles with task-scoped instructions - AGENTS.md § 10 — Model Selection Policy with cost guardrails

What to improve: Ensure docs/ai/usage-policy.md is linked prominently from onboarding materials and kept current as model selection policy evolves.

2. Healthy Data Ecosystem

AI is only as good as the data it can access. Poor-quality data (undocumented functions, missing type hints, silent errors) stymies AI effectiveness.

Fawkes implementation: - ruff + mypy enforce type hints on all Python services - Google-style docstrings required on all public functions - Structured logs (structlog) provide machine-readable audit data - BDD features in business language create readable acceptance criteria

What to improve: Track type-hint coverage as a metric in docs/METRICS.md. Target: 100% of public functions in services/ have type hints.

3. AI-Accessible Internal Data

Connecting AI to internal documentation, codebases, and decision logs turns a generic assistant into a specialised expert for your context.

Fawkes implementation: - docs/API_SURFACE.md — complete public interface map - docs/ARCHITECTURE.md — component relationships - docs/CHANGE_IMPACT_MAP.md — which files break when a component changes - AGENTS.md § 3 — ordered context file list agents must read first - RAG service (docs/ai/vector-database.md) for semantic doc search

What to improve: Ensure every new service has an entry in docs/API_SURFACE.md added in the same PR that creates the service.

4. Strong Version Control Practices

As AI accelerates code generation, robust version control is more critical than ever to maintain quality and enable safe experimentation.

Fawkes implementation: - Conventional commits: feat(scope):, fix(scope):, test(scope):, docs(scope): - PR size gate: > 400 lines blocked by CI; requires large-pr-approved label - Branch protection on main — no direct pushes - Every AI-generated commit must pass make lint before merge

Enforcement: AGENTS.md § 8 Instability Safeguards

5. Working in Small Batches

AI increases the risk of introducing many changes at once. Small-batch practices prevent speed from undermining stability.

Fawkes implementation: - PR size limit: 400 lines (CI gate) - Issue template requires explicit file list — agents must not create files not listed - AGENTS.md § 5 — agents must ask before touching more than 5 files in one task - DORA lead-time target: every issue completable in < 2 days of agent work

What to improve: Add a CI check that counts files changed and warns at 10+ files.

6. User-Centric Focus

Rapid AI-generated delivery only matters if aimed at real user needs. Teams must maintain tight feedback loops between delivery and user value.

Fawkes implementation: - BDD features written in business language (tests/bdd/) — user needs first - Backstage catalog (catalog-info.yaml) maps services to user-facing capabilities - Golden-path templates (templates/) reduce cognitive load for new engineers - Acceptance tests (AT-E1-*) validate platform capabilities from a user perspective

What to improve: Add user story links to every BDD feature file header.

7. Quality Internal Platforms

Internal platforms with paved paths, guardrails, and automated tooling are the critical multiplier for AI effectiveness. Poor platforms neutralise AI gains.

Fawkes implementation: - ArgoCD GitOps — declarative, automated reconciliation - Pre-commit hooks (make pre-commit-setup) — catch issues before CI - Helm charts with helm lint + helm template gates - make targets provide a single, documented CLI for all common tasks - Fawkes is the platform — every improvement here directly improves AI output quality

What to improve: Measure and report platform paved-path adoption (% of services using golden-path templates) in docs/METRICS.md.

AI Trust and Verify Protocol

Because ~30% of developers do not trust AI-generated code (DORA 2025), human review is a trust-building mechanism, not just a quality gate.

Read → Run → Review

1. READ   — AI reads the existing module before writing code.
            Never invent function names or import paths.

2. RUN    — AI executes tests after writing them.
            A test that has never run has unknown value.

3. REVIEW — Human approves all security, RBAC, secrets, and infra changes.
            Regardless of AI confidence level.

4. DECLARE — PR description notes which sections are AI-generated.

AI-Readiness Checklist

A module is "AI-ready" when agents can work on it without hallucinating context. Use this checklist before assigning an AI agent to a module:

  • [ ] Type hints on all public functions
  • [ ] Docstrings on all public classes and functions
  • [ ] Tests exist and are green
  • [ ] Module is single-purpose (not a God file)
  • [ ] Error messages include context (raise ValueError(f"createUser: {detail}"))
  • [ ] Module covered by at least one BDD scenario

Team Archetypes (DORA 2025)

DORA 2025 defines seven team archetypes. Fawkes targets the top tier:

Archetype AI Effectiveness Fawkes Target
Harmonious high-achievers Highest ✅ Target
Capable collaborators High Acceptable
Productive but siloed Medium Needs improvement
Legacy bottlenecks Low Unacceptable
Dysfunctional foundations Negative Blocked from AI use

Teams with "legacy bottlenecks" or "dysfunctional foundations" should fix their foundation before enabling AI tooling — AI will accelerate their problems.

Measuring AI Impact on DORA Metrics

Monitor these alongside classic DORA metrics when AI tooling is active:

Metric Positive Signal Warning Signal
Deployment frequency Increasing Unchanged or decreasing
Lead time for changes Decreasing Increasing (AI rework)
Change failure rate Stable or decreasing Increasing (AI-introduced bugs)
MTTR Stable or decreasing Increasing (harder to debug AI code)
Rework rate < 10% > 20% → stop, fix instructions
PR size Decreasing trend Increasing (AI batch risk)

Rework rate tracked in: docs/METRICS.md
Weekly check: scripts/weekly-metrics.sh