AI trends worth actually understanding

For most of the last three years, "using AI" meant sending one prompt to one model and reading one response. That era is over. The frontier has moved to systems of agents — multiple LLM-powered agents, each with a role, memory and set of tools, that plan, delegate, critique and hand off work to one another to complete tasks no single prompt could reliably finish.

A 2025 survey on LLM-based multi-agent systems frames this shift around a simple insight: a single model, however capable, is a bottleneck. It has one "voice," one context window and no built-in way to check its own mistakes. Multi-agent architectures fix this by splitting a task across specialised roles — a planner that breaks down the goal, workers that execute sub-tasks, and a critic or reviewer that checks the output before it ships.

The three shapes multi-agent systems take

Research on this topic (see the sources below) generally organises agent collaboration along a few key dimensions that are worth knowing cold if you're building or discussing these systems:

• Interaction type — cooperative (agents share a single goal), competitive (agents work against each other, e.g. adversarial testing), or "coopetitive" (a mix — common in negotiation and multi-stakeholder workflows).
• Structure — peer-to-peer (agents talk directly), centralized (one orchestrator agent routes all work), or distributed/hierarchical (teams of agents reporting up through layers, mirroring how a real org chart works).
• Coordination strategy — role-based (each agent has a fixed job, like "researcher" or "coder") or model-based (agents adapt their behaviour dynamically based on what the task needs).

Why this matters outside the research lab

These aren't academic distinctions. They map almost one-to-one onto the frameworks used in production today: LangGraph models workflows as explicit state machines (great for centralized, auditable orchestration); CrewAI leans into role-based cooperative teams (great for fast-to-build business workflows); AutoGen is built around conversational, peer-to-peer agent dialogue (great for open-ended problem-solving). Knowing why each framework exists — not just its syntax — is what separates someone who can debug an agent pipeline from someone who can only follow a tutorial.

Enterprises are adopting this pattern fast for exactly the reasons the research predicts: customer support triage that routes and escalates automatically, coding agents that write, test and fix their own output in a loop, and back-office workflow automation where one agent extracts data, another validates it against business rules, and a third takes the approved action. This is precisely the territory covered in our Gen AI & Agentic AI Frameworks and Multi-Agent Orchestration modules — the goal is that you leave able to choose the right architecture for a problem, not just wire up a demo.

Before November 2024, every team that wanted an AI agent to use a tool — query a database, call an internal API, browse a file system — wrote custom "glue code" for that specific model and that specific tool. Connect 5 models to 10 tools and you needed up to 50 separate integrations. Anthropic's Model Context Protocol (MCP) proposed a fix: one open, standardised protocol that any model can speak to any tool through, the same way USB-C lets any device talk to any charger.

What happened next is one of the fastest standardisation stories in software history. Within its first year, MCP reached roughly 97 million monthly SDK downloads and 10,000+ active MCP servers in the wild — and, notably, it wasn't just Anthropic's own models using it. OpenAI, Google and Microsoft all adopted MCP for their own agent platforms, an unusual show of cross-competitor agreement that only happens when a technical standard genuinely solves a shared problem.

From one company's protocol to neutral infrastructure

The clearest signal that MCP has "won" came in December 2025, when Anthropic donated the protocol to the newly formed Agentic AI Foundation, a directed fund under the Linux Foundation — co-founded by Anthropic, Block and OpenAI. Handing a protocol to neutral, vendor-independent governance is exactly what happened to HTTP, USB and Kubernetes on their way to becoming permanent infrastructure rather than one company's product. It signals that the industry expects MCP to outlast any single AI lab's product roadmap.

What an MCP server actually does

Structurally, MCP defines three roles: a host (the AI application, e.g. an IDE or chat client), a client (the connector living inside the host), and a server (a lightweight program that exposes a specific tool, database or API in the standard MCP format). Once you've built one MCP server for, say, your company's CRM, any MCP-compatible model or agent framework can use it immediately — no bespoke integration required.

For students, the practical takeaway is that "I can build an MCP server" is quickly becoming as fundamental a résumé line as "I know REST APIs" was a decade ago. It's why MCP gets its own dedicated module in our Gen AI & Agentic AI program — including hands-on server-building — and why it shows up again in the SAP AI with BTP course, where SAP's own Agent Builder and Joule ecosystem increasingly lean on the same tool-connection philosophy.

LLMs are trained on a fixed snapshot of the internet and have no idea what's inside your company's contracts, tickets or product database. Retrieval-Augmented Generation (RAG) closes that gap: instead of asking a model to answer from memory, you first retrieve relevant documents from your own data, then hand them to the model as grounded context. It remains, by a wide margin, the most-used pattern for putting generative AI safely in front of real business data.

But "RAG" in 2026 looks very different from the naive "embed, retrieve top-5, stuff into prompt" pipelines of 2023. A comprehensive 2025 survey of RAG architectures traces this evolution across three generations: naive RAG (simple similarity search), advanced RAG (adding re-ranking, query rewriting and hybrid keyword+semantic search to fix weak retrieval), and modular / agentic RAG, where retrieval itself becomes a multi-step reasoning process the model actively controls.

What makes RAG "agentic"

A dedicated survey on Agentic RAG describes the key shift clearly: rather than retrieving once and generating once, an agent can now decide whether to retrieve at all, issue multiple rounds of retrieval for multi-hop questions ("first find the product, then find its warranty policy"), rewrite a poorly-phrased query before searching, and critique its own draft answer against the retrieved sources before returning it. This turns RAG from a static pipeline into something closer to how a good research analyst actually works.

• Hybrid search — combining keyword search (great for exact terms like product SKUs) with semantic vector search (great for meaning) outperforms either alone.
• Re-ranking — a second, more expensive model re-scores the top retrieved chunks, since the fastest first-pass retriever isn't always the most accurate one.
• Groundedness evaluation — measuring whether every claim in the generated answer is actually supported by the retrieved text, which is now a standard interview and production-monitoring metric.

The enterprise angle

This matters especially for enterprise AI, where the retrieved "documents" aren't public web pages — they're proprietary, permissioned, and often scattered across systems like SAP S/4HANA, SharePoint or a data warehouse. That's exactly the problem SAP's Generative AI Hub, Datasphere and Business Data Cloud are built to solve on the enterprise side, and it's why RAG gets deep, hands-on coverage in both our Gen AI & Agentic AI program (its own dedicated module) and the AI & Data Science program's Generative AI unit — because whether you end up building consumer AI products or enterprise copilots, grounding output in real data is the single skill that most determines whether an AI feature is trustworthy enough to ship.

Say "foundation model" to most people and they picture ChatGPT or an image generator. But most real-world data science work isn't text or images at all — it's spreadsheets: customer churn tables, transaction logs, sensor readings, medical records. Rows and columns. And until recently, the best tool for that kind of data was still what it had been for a decade: carefully hand-tuned gradient boosting (XGBoost, LightGBM) or Random Forests, rebuilt from scratch for every new dataset.

That changed in January 2025, when a team led by Prof. Frank Hutter at the University of Freiburg published TabPFN in Nature — a transformer-based foundation model built specifically for tabular prediction. TabPFN is a "Prior-Data Fitted Network": instead of training on your dataset the way XGBoost does, it was pre-trained once, on millions of synthetic datasets generated from a broad prior over plausible causal structures. When you show it a new, real dataset, it doesn't retrain at all — it does a single forward pass and produces predictions that approximate proper Bayesian inference on the fly.

Why this is a genuinely big deal, not just hype

In the original paper's benchmarks, TabPFN matched or beat carefully tuned gradient-boosted trees on small-to-medium tabular datasets (under roughly 10,000 rows) — datasets that describe the vast majority of real business problems, where you rarely have Kaggle-scale millions of rows. And it does this in seconds, with no hyperparameter search, no cross-validation loop, no GPU cluster.

The follow-up, TabPFN-2.5 (released November 2025), pushed the model's capacity roughly 20x further — handling datasets up to 50,000 rows and 2,000 columns — closing much of the gap with larger enterprise datasets that the original version couldn't comfortably handle.

• What it changes day-to-day: for small/medium datasets, a first-pass model can now take minutes, not days — freeing up time for feature engineering and problem framing, the parts of data science that actually require human judgement.
• What it doesn't change: for very large-scale or highly domain-specific problems, tuned gradient boosting and deep learning are still very much in play — TabPFN is a powerful new tool in the kit, not a replacement for understanding the fundamentals.
• Why that balance matters for interviews: being able to explain when to reach for a tabular foundation model versus classical ML — not just that one exists — is what signals real understanding to a hiring manager.

This is exactly the judgement our AI & Data Science program is built to develop: the curriculum grounds you thoroughly in Scikit-learn, XGBoost and LightGBM first, so that when a tool like TabPFN comes up — in this course or the next one after it — you can evaluate it critically instead of chasing whatever's trending.