Ask a frontier model to design your infrastructure and you'll get something plausible. Plausible is the problem. It doesn't know that you already run two EKS clusters, that your platform team banned serverless last quarter, or that everything in production must come from your approved catalog. You don't need plausible infrastructure. You need yours.
Today we're introducing the Architect Agent: describe what you're building in plain language, and get back a reviewable architecture design that is grounded in your real environment, constrained by your organization's rules, and ready to deploy through DevPlus's governed pipeline.
Research First, Design Second
Before the Architect proposes anything, it is required to gather evidence. It pulls a deterministic digest of your discovered cloud environment (resource counts, container platforms, networks, data stores, internet exposure), computed by code, not by a model. It searches a curated corpus of architecture decision playbooks (ECS vs. EKS, ALB vs. NLB, VPC topology, and more, plus any playbooks your team uploads). And it reads your catalog to know which building blocks actually exist.
The result is a design where every component and every major choice carries citations: the knowledge excerpt that informed it and the environment fact that justified it. "You already run two EKS clusters in us-east-1" is an argument for reusing an orchestrator, and the design says so, explicitly.
Architecture Principles: Your Opinions, Enforced
This is the part we care about most. Admins author Architecture Principles: directives like "microservices deploy to the shared cluster," deny-lists like "no serverless," or a catalog-only rule that restricts designs to approved building blocks. Principles are injected into every single design run. They are not retrieved when relevant; they are always there.
Catalog-only and deny rules are enforced by a deterministic validator, not by asking the model nicely, while directives are injected into every design as non-negotiable requirements. A design that breaks a hard, checkable principle is never even presented for approval; the run fails closed with an explanation. Advisory principles surface as warnings. And every principle that shaped a design is cited on the review card, so your team can see the rules working.
Any LLM can suggest infrastructure. DevPlus enforces yours.
Gaps Are Decisions, Not Dead Ends
Sometimes a design needs a building block your catalog doesn't have. The review card turns each gap into a decision: provide an existing catalog item, skip the component, or have the Architect author it: generate the Terraform module, create a private GitHub repo through your GitHub App connection, validate it with tofu validate, publish it to your private module registry, and wrap it as a catalog item. Every file is shown on a review card before anything is written, and validation failures drive a bounded, human-gated self-correction loop.
Approve, Then Deploy in Waves
An approved design executes component by component, in dependency order, through the exact same pipeline as any manual deployment: OpenTofu plan, cost estimate, Cedar policy gate, then apply, each run in its own isolated Kubernetes job pod. Outputs from completed components flow into the inputs of the ones that depend on them. If a dependency is still provisioning, the run pauses honestly and picks up the next wave when you continue, across page reloads, restarts, even days.
Built Like We Don't Trust It
The Architect holds no cloud credentials and no GitHub credentials. Every side effect goes through the backend under your own session, authorized by the same Cedar policies that govern humans. Every write pauses on an approval card, and the exact payload you approve is hashed into a durable audit log. Edited designs are re-validated with the same deterministic rules before execution. If your principles can't be loaded, no design is produced at all.
That's the shape of AI we believe in for infrastructure: intelligent where it helps, deterministic where it matters. The Architect Agent is available now: connect a cloud, run discovery, write down your principles, and ask it for a design.