AI stack trends across 7 companies: what The Pragmatic Engineer actually shows

The Pragmatic Engineer published a snapshot of the AI stack at seven companies. Different businesses, different profiles — but three very recognizable patterns:

• AWS Bedrock as the preferred way to run Anthropic models.
• Postgres with pgvector as the default for embeddings and vector search. Wordsmith is the lone exception, using Pinecone.
• LangChain showing up as the LLM integration layer at several places.

And one observation that stuck with me more than the rest:

The bigger the scale, the closer you get to the "metal." Augment Code running on NVIDIA GPUs with CUDA is the clear example.

The obvious takeaway is "note down those tools." The useful one is different: these choices trace a curve, and figuring out where your team is on that curve resolves most stack decisions before you even open a comparison.

Bedrock as the default: managed wins early

It makes sense that most teams start on Bedrock to serve Claude.

You inherit IAM, VPC, observability, consolidated billing, and a predictable compliance path. For a team validating whether an agent solves a real business problem, those guarantees are worth more than shaving a few cents per million tokens.

The Bedrock choice is rarely about the model. It's about not creating new operational surface at a moment when the rest of the system is already moving fast.

pgvector wins the early phase — and that's not laziness

Pinecone, Weaviate, Qdrant, Milvus, all have real merits. Even so, six of the seven companies picked pgvector.

The reason is less about performance and more about cognitive surface:

• there's already a Postgres in the company;
• there's already a runbook, backups, replication, metrics;
• embeddings sit next to the data they describe;
• joins between vectors and metadata are trivial.

Trading that for a dedicated vector database only pays off when you hit a latency or index-size limit that pgvector + hnsw can't handle. Before that point, it's premature optimization wearing an architecture-decision costume.

Wordsmith landing on Pinecone doesn't break the pattern — it probably is the pattern: they hit a wall and migrated. The rule still holds: start on pgvector, migrate when it hurts.

LangChain: the layer nobody loves, everybody uses

LangChain became a meme in some corners, but it keeps showing up in these stacks. Why?

Because the problem it solves is real: chaining prompts, tools, retrievers, memory, and provider fallbacks without reinventing the same glue every week. If you're alone on a side project, you can write that by hand. At a company with five engineers shipping AI features, somebody is going to need the abstraction — and LangChain is there.

What changed over the past year:

• usage got more surgical (orchestration, not everything);
• many teams keep their own wrappers around it to isolate API churn;
• alternatives (LlamaIndex, Vercel AI SDK, agent-specific frameworks) split the territory.

It's still a reasonable starting point. Just don't treat it as a permanent contract.

The inflection: when you go down to the metal

Augment Code is the interesting case in the snapshot. They're not on Bedrock — they're renting NVIDIA GPUs and writing CUDA.

This isn't engineering vanity. It's a direct response to two pressures:

1. Constant fine-tuning: you need control over the training pipeline, not just inference.
2. Inference volume: at their scale, the gap between "tokens via API" and "tokens via your own GPU" becomes the product.

When those two pressures show up together, managed stops being cheap. But if one of them isn't on your plate, going down to the metal just buys you a year of problems you didn't need.

The curve, in one sentence

The Pragmatic Engineer snapshot describes a trajectory that keeps repeating:

Managed as long as possible, dedicated when the cost of not doing it grows past the cost of doing it.

Translated into concrete decisions:

• Start on Bedrock (or equivalent) with pgvector and some orchestration framework.
• Put a thin wrapper between your domain and any provider — you'll swap models more often than you think.
• Only consider a dedicated vector DB after EXPLAIN ANALYZE becomes the problem, not before.
• Only consider your own GPUs when fine-tuning and inference volume are genuinely a competitive edge.

The right question isn't "which stack should I pick?" It's "where am I on the curve right now, and what's the next trigger that moves me?".

When you answer that one honestly, most tooling decisions resolve themselves.

---

Source: The Pragmatic Engineer — Tech stack trends across companies.