Is Spec-driven Development Key For Infrastructure Automation?

Since GitHub Universe and nan announcement of GitHub Spec Kit, spec-driven improvement (SDD) has taken nan dev world by storm. The premise is compelling: Give AI agents system discourse done markdown specs earlier they constitute code, and you’ll person it all. An extremity to hallucinations astir APIs, rushed coding and low-quality outcomes.

With SDD, AI agents will activity much for illustration quality developers who person merchandise requirements documents (PRDs), break down tasks and execute systematically.

The conception formalizes what improvement teams person done for years. A merchandise head writes requirements. Developers digest nan PRD aliases specifications, break nan activity into tasks and commencement coding. SDD simply structures this workflow for nan AI era, turning earthy connection specifications into nan discourse large connection models (LLMs) request to make meaningful code.

As personification who lives and breathes DevOps and level engineering, I recovered myself asking nan evident question: What does this mean for infrastructure work? Should we beryllium racing to adopt SDD for our Terraform modules and Kubernetes configurations?

Infrastructure Code Isn’t Application Code

Infrastructure codification looks for illustration code, but it behaves very otherwise from exertion code.

Look astatine immoderate Terraform file, Helm floor plan aliases CloudFormation template. What do you see? Specifications. Infrastructure arsenic Code (IAC) is already spec-driven by design. It is declarative. We picture nan desired state. We opportunity “I want a database pinch these properties,” not “Execute these commands to create a database.”

But here’s wherever things get interesting.

• Application codification favors creativity. Give 10 developers nan aforesaid characteristic request, and you’ll get 10 different implementations. Each mightiness beryllium valid, elegant successful its ain way. The extremity is to lick business problems pinch caller approaches, optimize for personification acquisition aliases find clever capacity improvements. There’s worth successful that diverseness of solutions.
• Infrastructure codification favors reproducibility. When I rotation up infrastructure successful us-east-1, eu-west-1 and ap-southeast-1, I request identical configurations. Same networking setup, aforesaid information groups, aforesaid database configurations. Standardization intends predictable costs, interchangeable parts and reliable disaster recovery.

This favoritism matters for SDD because AI agents thrive connected imaginative problem-solving but struggle pinch strict reproducibility. We don’t want an AI supplier getting imaginative pinch our virtual backstage unreality (VPC) configuration. We want nan nonstop aforesaid blueprint deployed perfectly each time.

More importantly, infrastructure codification seldom flows from spec to implementation. Consider really infrastructure really evolves. FinOps adjusts lawsuit types to optimize costs. Security patches a vulnerability straight successful production. Someone scales a database done nan console during an incident. Your Terraform still describes nan original state, but reality has moved on. This is drift: erstwhile your existent unreality resources nary longer lucifer your IaC. Infrastructure teams activity backward, perpetually updating specifications to lucifer reality.

​​SDD assumes a guardant travel from requirements to code. But that’s not nan measurement level teams work. We don’t request AI to constitute much Terraform from specs; we request thing other entirely.

The Real Automation Gap Is Deployment Orchestration

SDD will power infrastructure work, but not ever successful ways level teams will celebrate.

Today, developers are already 10 times much productive pinch AI copilots than they were conscionable a fewer years ago. SDD promises to push this moreover further: complete modules generated from specifications, full features materialized from markdown plans. The measurement of exertion codification will explode.

All this codification needs location to run. Every characteristic needs infrastructure. Every microservice needs its database, connection queue and networking. The acceleration successful codification accumulation creates unprecedented unit connected deployment velocity.

Yet deployment remains stubbornly manual. While developers get AI assistants that move specs into code, level engineers still coordinate analyzable deployments by hand. We tin make a complete microservice successful a fewer hours, but walk days figuring retired really to safely group up and deploy its infrastructure.

Why AI Agents Can’t Orchestrate Infrastructure Deployment

The barriers to AI-driven deployment are structural problems successful nan measurement infrastructure codification is organized today.

• Terraliths: Monolithic nightmares. We’ve created monolithic Terraform files wherever specifications, values and logic are tangled together. A azygous record mightiness specify networking, databases and exertion configuration each mixed up. Small changes cascade unpredictably. There’s nary measurement to understand blast radius erstwhile everything touches everything else.
• Heterogeneous tooling. A azygous situation uses Terraform for infrastructure, Helm for Kubernetes, arsenic good arsenic Python scripts. Each pinch different inputs, outputs and authorities management. Orchestrating crossed them requires knowing hidden limitations that aren’t documented anywhere.
• Working backward from drift. Infrastructure teams perpetually reconcile drift, updating codification to lucifer reality. AI agents would first need to representation what really exists crossed each clouds, regions and accounts earlier moreover attempting to update anything.

What we really request is to restructure infrastructure to beryllium AI-ready.

Blueprint-Driven Deployment: Infrastructure for nan AI Era

The way guardant is clear. We request to toggle shape nan measurement infrastructure is packaged, deployed and managed. Here’s really to make infrastructure AI-ready:

• Transform each portion of infrastructure. Turn each Terraform module, Helm floor plan and Python book into artifacts pinch normalized inputs and outputs. These artifacts go reusable building blocks that combine into blueprints.
• Assemble artifacts into clear, versioned blueprints with well-defined boundaries. A database blueprint creates databases. A networking blueprint handles VPCs and subnets. No mixing, nary confusion.
• Publish them to a catalog. Decide which ones to aboveground to AI agents truthful they cognize what’s safe and allowed to deploy.

This attack solves nan structural problems we mentioned earlier:

• Terraliths get decomposed into artifacts and blueprints. That 10,000-line monolith becomes a postulation of focused components, each pinch its ain life rhythm and clear interfaces. Changes are scoped. Blast radius is contained. AI useful pinch nan blueprints, not nan tangled code.
• Heterogeneous tooling becomes unified. Terraform, Helm, Python, etc., each go artifacts pinch modular inputs and outputs done normalization. The orchestration furniture doesn’t attraction what created nan artifact.

With that successful place, you tin tackle nan drift problem:

1. Regularly representation what exists: Discover existent unreality authorities crossed each accounts and regions into a unreality graph.
2. Extract patterns from accumulation and create/update blueprints.

The rhythm becomes sustainable: Reality informs blueprints, blueprints guideline deployment and deployment is tracked successful nan graph.

Now an AI supplier tin really work:

Request: “We request to grow to Asia-Pacific for little latency.”

The supplier queries nan blueprint catalog and unreality graph, and finds nan modular location infrastructure blueprint already moving successful nan United States and Europe. It understands limitations from nan graph, networking databases earlier applications.

“I’ll deploy location infrastructure blueprint v2.3 to ap-southeast-1. Based connected nan dependency graph, this takes 12 minutes. No existing resources affected.”

One approval. The orchestrator handles nan rest.

Infrastructure Is Different From Application Code

Spec-driven improvement makes consciousness erstwhile you’re building features. It doesn’t erstwhile you’re building nan level those features tally on.

Infrastructure needs:

• Blueprints, not specs: Reusable, versioned patterns you tin rotation retired crossed regions and environments.
• Orchestration, not conscionable codification generation: Coordinated, multistep workflows crossed infrastructure, configuration and apps.
• Clear boundaries, not entangled modules: Well-defined scopes per blueprint and artifact truthful changes enactment contained.
• Cloud graphs, not codification dumps: A unrecorded position of what really runs successful each account, region and cloud.

The early isn’t AI agents generating Terraform. It’s AI agents executing deployments safely by utilizing prevalidated blueprints.

This is really AI yet enters nan deployment game.

Group Created pinch Sketch.