A working map for practicing architects: where AI, multi-cloud, platforms, zero trust,
data and events, FinOps, and sustainability show up in real designs—and how to keep
decisions tied to shipping systems, not slide decks.
The job is less “draw the system” and more “make sure the system keeps its promises.”
You are trading off models, regions, data residency, agent workflows, and platform
defaults every week. What follows is an operating model, a reference picture of the
stack, decision habits that survive contact with engineering, and KPIs you can actually
track (DORA, SPACE, FinOps-style economics—used as shorthand, not as doctrine).
5
Strategic pillars
12
Architectural capabilities
7
Decision frameworks
24
KPIs & fitness functions
02Strategic Framework — The Five Pillars
I group the work into five pillars—business alignment, AI-native design, platform
delivery, security and trust, FinOps and sustainability. Each has its own principles
and a few measurable checks; the point is to reuse the same mental model whether you are
on one product team or across a federated org.
Figure 1 — Five strategic pillars of the modern Solution Architect operating model.
P1 · Business Alignment
Map systems to value, not technology
Every solution must trace to a measurable business capability and a strategic
north-star metric. Wardley maps surface evolutionary stages; capability heatmaps
expose duplication and gaps.
P2 · AI-Native Design
Treat models as first-class components
LLMs, retrieval, vector stores, and agents are now load-bearing primitives. Design
for evaluations, drift, hallucination budgets, and human-in-the-loop checkpoints from
day one.
P3 · Platform & Delivery
Reduce cognitive load, increase flow
Architects shape the Internal Developer Platform so product teams ship through
paved roads. Treat the platform as a product with users, SLOs, and adoption metrics.
P4 · Security & Trust
Assume breach, prove integrity
Zero trust, SLSA-attested supply chains, signed artifacts, and confidential
compute are baseline. Architectural decisions encode regulatory posture (DORA-EU,
AI Act, NIS2, GDPR).
P5 · Sustainability & FinOps
Design for cost and carbon
Unit economics and gCO₂e per request are now non-functional requirements.
Carbon-aware scheduling, right-sizing, and tier-based storage become design inputs.
Cross-cutting
Evolutionary by construction
Architecture is a continuous flow, not a milestone. Fitness functions, ADRs, and
architectural intent in code keep the system honest as it evolves.
03Role Model — What a Solution Architect Owns
Most solution architects work at four altitudes at once—enterprise,
domain, solution, and component. Naming the altitude saves you from living only in
ivory-tower slides or only in someone else’s pull request.
Figure 2 — Four altitudes the Solution Architect operates across. Each carries its own horizon, stakeholders, and artefacts.
Core deliverables by altitude
Altitude
Primary Decisions
Key Artefacts
Cadence
Enterprise
Build vs. buy vs. compose; sovereign vs. hyperscaler; reference standards
API contracts · schemas · concurrency · failure modes
OpenAPI / AsyncAPI · ADRs · architectural unit tests
Per sprint
04Reference Target Architecture (2026)
Below is a vendor-neutral reference stack: horizontal planes (experience through
foundation) with security, observability, FinOps, and governance called out as
cross-cutting. Adapt labels to your estate; the layering is the takeaway.
Figure 3 — Vendor-neutral reference architecture organised into six capability planes with cross-cutting concerns.
05Decision Frameworks & Architectural Methods
Most of your leverage is in explicit decisions—what you chose, what you ruled out, and
what would make you change your mind. The toolkit below mixes lightweight habits (ADRs,
C4) with heavier enterprise methods where they earn their keep.
Decision Records
ADR (Architecture Decision Records)
Markdown-versioned, in-repo records of every significant decision: context,
options, decision, consequences. Linked from C4 diagrams.
Visualisation
C4 Model + arc42
C4 for layered diagrams (Context → Container → Component → Code); arc42 as a
structured documentation template. Diagrams-as-code via Structurizr / Mermaid.
Strategy mapping
Wardley Mapping
Position components on a value-chain × evolution axis to surface what to build,
buy, outsource, or retire. Drives platform vs. product investment.
Trade-off analysis
Quality Attribute Workshop (QAW) & ATAM
Elicit and prioritise NFRs, identify sensitivity and tradeoff points across
quality attributes. Outputs feed risk register and fitness functions.
Enterprise
TOGAF ADM & capability mapping
Use TOGAF's phases as a checklist, not a religion. Capability maps remain the
most durable enterprise artefact for prioritisation.
Evolutionary
Fitness Functions
Automated tests in pipelines that enforce architectural characteristics — latency,
dependency direction, coupling, license, cost-per-request — turning intent into code.
Architect's decision flow
Figure 4 — Architect's decision flow. Every step produces an auditable artefact and feeds the next.
06Non-Functional Requirements & Fitness Functions
Where you can, turn quality attributes into checks in the pipeline—latency
budgets, coupling rules, cost per request, whatever your team will actually run. The
table is a starting template per solution, not a standard you must paste wholesale.
DORA metrics · ArchUnit-style coupling tests · test pyramid
07AI-Native & Agentic Architecture Patterns
If models sit on the critical path, they get the same treatment as payments or auth:
SLOs, evals, cost envelopes, and kill switches—not a sidebar “AI workstream.”
Pattern
LLM Gateway with model routing
Centralised gateway routes by capability/cost/latency, applies rate limits, PII
redaction, prompt caching, and structured logging. Decouples app code from model
vendor; enables A/B and shadow evaluation.
Tool-using agents over Model Context Protocol servers, with explicit planning,
budgets, sandboxes, and trace-replayable runs. Human-in-the-loop on irreversible
actions.
Pattern
Evaluation-driven development
Golden datasets, LLM-as-judge with calibration, regression suites in CI, and
online quality SLOs. No model change ships without eval delta.
Pattern
Responsible-AI guardrails
Input/output classifiers, policy engines, content provenance (C2PA), and audit
logs aligned to the EU AI Act risk tiers and ISO/IEC 42001.
Pattern
Small + specialised models
Use frontier models for reasoning; distil or fine-tune small models for hot paths.
Quantise for edge. Result: lower latency, lower cost, lower carbon.
08Security, Trust & Regulatory Posture
EU AI Act, DORA (where it applies), NIS2, GDPR, and sovereign-cloud expectations show up
in landing zones, data flows, and control design—waiting for a compliance review to
“sign off” the architecture is how you paint yourself into a corner.
Architectural rule: regulatory obligations are encoded as
policy-as-code at the platform layer, so product teams inherit compliance by
default rather than re-implementing it per service.
Crypto-agility, hybrid TLS
(X25519 + ML-KEM), inventory of long-lived secrets and certificates.
0912-Month Adoption Roadmap
Ship this as a quarterly plan: theme, a few measurable outcomes, and clear “we are done”
lines. Architecture should be co-owned with engineering and product—you are not running
a parallel PMO.
Figure 5 — Quarterly adoption roadmap. Each quarter delivers usable platform capabilities, not slideware.
10Measuring Architectural Success
If you cannot point to numbers, you will lose arguments to urgency. Mix delivery,
reliability, developer experience, and unit economics—the exact targets below are
examples; swap in what your org already reports.
Category
Metric
Target / Direction
Delivery (DORA)
Deployment frequency
Multiple / day
Lead time for change
< 1 day
Change failure rate
< 10%
MTTR
< 1 hour
Developer experience (SPACE)
Time to first commit
< 1 day
Cognitive-load score
↓ each quarter
Golden-path adoption
> 80%
Reliability
SLO attainment
≥ 99.9% of SLOs green
Error-budget burn
Within policy
Economics
Cost per business transaction
↓ trending
Carbon per transaction
↓ trending
AI quality
Groundedness score
≥ 0.85
Harmful output rate
< 1%
11Anti-Patterns to Avoid
Anti-pattern
Ivory-tower architecture
PDFs and diagrams disconnected from running code. Counter: diagrams-as-code,
ADRs in the repo, fitness functions in CI.
Anti-pattern
Big design up front
Locking decisions before learning. Counter: distinguish one-way vs. two-way doors;
defer reversible decisions to the last responsible moment.
Anti-pattern
AI as a feature flag
Bolting an LLM onto a UI without evals, guardrails, or cost controls. Counter:
treat AI as a critical-path subsystem with its own SLOs.
Anti-pattern
Platform without users
Building an IDP nobody adopts. Counter: platform-as-product with PMs, SLOs,
adoption metrics, and feedback loops.
Anti-pattern
Cloud lift-and-shift theatre
Re-hosting VMs with no architectural change. Counter: re-platform around managed
services, events, and serverless where appropriate.
Anti-pattern
Single-vendor lock-in by default
Counter: provider-neutral abstractions for compute, identity, and data plane;
measured exception process for managed services.
12Research Agenda & Open Questions
The list below is deliberately unfinished—places where I still want sharper methods and
better data, not polished conclusions.
Agentic system safety — provable bounds on tool-use blast radius
and economic guardrails for autonomous agents.
Continuous architecture — formalising fitness functions as the
primary architectural contract instead of static diagrams.
AI-assisted architecting — copilots that propose ADRs, detect
drift, and synthesise C4 views from running systems.
Sustainable software — standard methodologies for measuring
software carbon intensity (SCI) at the request level.
Post-quantum migration — pragmatic roadmaps for crypto-agility in
long-lived enterprise systems.
Sovereign multi-cloud — workload portability under conflicting
jurisdictional requirements.
13Key Concepts (Glossary)
TOGAF 10arc42C4 ModelWardley MappingDDDTeam TopologiesDORASPACESLSASPIFFE / SPIREOpenTelemetryCloudEventsAsyncAPIOPA · CedarSigstoreEU AI ActDORA-EUNIS2ISO/IEC 42001FinOps FrameworkGreen Software Foundation SCIMCPC2PA
Shorthand I use in workshops and reviews so the same words mean the same thing across
business, engineering, security, and regulators. For citable sources and specifications,
see §14 · References.
14References
Figures, narrative, and synthesis in this playbook are original work by the author.
The list below grounds methods, metrics, regulations, and platform primitives in
primary documents—standards, authoritative books, and official specifications. It is
not exhaustive; use it as a trailhead for procurement, security review, and audit.
Enterprise architecture, documentation, and decision records