PulseFlow Tecnologia
← Insights
·8 min read·market-analysis·ai-stack·mcp·langgraph
PulseFlow Tecnologia

The new enterprise AI stack

In 2024, the AI agent stack fit into a few obvious layers. The 2026 version, mapped by O'Reilly, has six distinct layers - and at least three of them simply didn't exist as a separate category two years ago. That alone says something about how fast agent infrastructure matured.

What changed between 2024 and 2026

Three forces redrew the map. MCP standardized tool connectivity, creating an entirely new "Protocols and Tools" layer - now at 97 million monthly SDK downloads and adopted by OpenAI, Google, and Microsoft. Reasoning models (o1, o3, DeepSeek R1, Claude with extended thinking) let agents solve problems in a single call, eliminating multistep chains that used to be mandatory. And memory stopped being "a subset of a vector database" and became a first-class architectural primitive - "context engineering" itself replaced "prompt engineering" as the central discipline.

The six layers, each with its own risks

Models and inference: open-weight models (Llama 3.3, DeepSeek V3, Qwen 2.5) have already closed much of the quality gap with closed models, creating a common pattern of prototyping closed and deploying on open-weight.

Protocols and tools: the same layer MCP made possible also brought new risk - an Endor Labs analysis of 2,614 MCP servers found 82% prone to path traversal and 67% to code injection.

Memory and knowledge: with three tiers - in-context state (windows up to 1M+ tokens in Gemini), vector search (pgvector as the default), and persistent cross-session memory (GraphRAG) - this is the highest state-complexity layer in the stack, and also the one that breaks most often in the prototype-to-production transition.

Frameworks and SDKs: LangGraph (v1.0 in October 2025) already runs in production at companies like Uber, JPMorgan, LinkedIn, and Klarna - but this is the highest lock-in-risk layer in the whole stack, because orchestration code isn't portable between frameworks.

Evaluation and observability: the most revealing data point in the whole report comes from LangChain's State of Agent Engineering (December 2025): 89% of teams with agents in production have already implemented observability, but only 52% have implemented evals - a 37-point gap exactly where production quality fails silently.

Guardrails and security: the least mature layer in the stack, with no dominant framework. The pattern that emerged is "guardrails before the action" - authorization at the tool-execution layer, not at the output, because filtering the response after the email has already been sent doesn't help.

The stack will collapse, according to O'Reilly itself

The report's conclusion is direct: "the stack will collapse" - provider SDKs are already absorbing memory, tool calling, and basic eval into a single API. The prediction is that, by 2027, most teams will have an increasingly opinionated stack from their own model provider, sufficient for 80% of use cases - while the remaining 20%, agents at scale where standards break, will keep requiring layer-by-layer customization.

Sources