Steve Hedden in his recent article Is RAG Dead? The Rise of Context Engineering and Semantic Layers for Agentic AI argues about how enterprise AI is evolving from static piepeline RAG (Retrieval-Augmented Generation) into a system which is delivering context engineering and semantic layering, that govern what information the system pulls in, under what rules, and with what traceability.

Julian Schrittwieser in his article Failing to Understand the Exponential, argues about tempo: we repeatedly underestimate the speed at which capabilities improve, which makes any fixed “best model” choice fragile.

Read together, the message is simple but demanding: build systems that are deeply grounded in your own context, yet easy to upgrade as capability jumps.

Hedden’s thesis, unpacked (beyond “RAG is dead”)

Hedden’s headline critique isn’t really about RAG; it’s about the shallowness of naïve retrieval. The failure mode he targets is familiar: throw documents into a vector index, retrieve a few similar chunks, and hope the model says something useful. That approach breaks the moment you need business guarantees.

The alternative is context engineering supported by a semantic layer:

From text to meaning. Instead of treating “whatever is semantically similar” as good enough, you declare what entities actually are in your business (customers, products, policies, claims) and how they relate. That ontology can live in a knowledge graph or a product that approximates one, but the essence is typed, connected information, not loose paragraphs.
Governance where retrieval happens. Hedden’s emphasis on policy-as-code matters: role and data rules should be enforced when the system assembles context, not bolted on after the answer is generated. That’s how you avoid “the model said something it should never have known” incidents.
Provenance as a first-class feature. Answers must carry traceable citations - to datasets, tables, document sections, versions, and dates. This isn’t academic fussiness; it’s how teams debug, audit and learn. Without provenance, you can’t improve the system except by superstition.
Tools and memory under discipline. Agents don’t just read; they act. The semantic layer decides which tool to call (SQL, API, calculator), what minimal context to pass (for speed and cost), and when to escalate (ask a human, switch model, request permission). That keeps behaviour predictable.

What Hedden is not saying is that retrieval is obsolete. He’s saying retrieval must be governed by meaning, rules, and receipts. In other words, less “search the heap,” more “assemble the case”.

Schrittwieser’s thesis, unpacked (beyond “AI is exponential”)

Schrittwieser’s point is sharper than generic hype: organisations consistently under-model the slope of capability improvement. In particular, he highlights evidence that the length of tasks AI can complete autonomously (at ~50% success) is doubling roughly every seven months (see the METR series he cites below). That cadence has three practical consequences:

Model behaviour will change more often than your planning cycle. Windows get bigger, tool use improves, reasoning gets more robust, and price/performance shifts. Treating any given model as a fixed asset is an error of horizon.
Benchmarks age quickly. A model that looked second-tier last quarter may be “good enough” for most tasks this quarter, often at lower cost and higher throughput.
Architectures must absorb swaps. If your pipeline hard-codes prompt formats, tool contracts, or context assembly to one vendor’s quirks, you turn the slope into a cost centre.

What Schrittwieser is not saying is that you should chase every shiny update. He’s saying design for churn so you can selectively take the upgrades that matter.

How these two arguments reinforce each other

In essence Hedden gives you reliability and accuracy while Schrittwieser keeps you adaptable. A governed semantic layer ensures answers are grounded and compliant regardless of which model sits underneath. That makes model swaps far less risky.
Schrittwieser explains the business value of Hedden’s discipline. If we accept faster-than-expected capability jumps, then the only sane hedge is to decouple business logic and data meaning from model choice. That is precisely what context engineering achieves.
Both reject “bigger is automatically better.” Hedden redirects attention to better context; Schrittwieser suggests many tasks will be served by smaller, cheaper models, reserving heavy models for genuinely hard or risky work. Together, that’s a system-level, not model-centric, view of performance.

Implications for companies investing in digital transformation

1) Prioritise meaning and rules before model selection

If your data is inconsistent (different names for the same thing, unclear “system of record”) or if access rules only live in app front ends, changing models won’t help. Invest in the semantic spine first: shared definitions, clear ownership of truth, and rules enforced at retrieval. It makes every future capability jump safer to adopt and more valuable once adopted.

2) Treat provenance like uptime

Make “show your working” a default behaviour, not an afterthought. Citations are how people verify, how auditors sign off, and how engineers debug. Without provenance, improvements are guesswork and trust never compounds. You won’t scale usage internally (or expose AI to customers) if no one believes the answers.

3) Make model swapping a design constraint

Assume you’ll have to change, combine, or re-route models as prices, latencies, and behaviours shift. Keep prompts, retrieval recipes, tool contracts, and evaluation harnesses modular so they can evolve without a rewrite. You capture upside from the slope without serial re-platforming.

4) Optimise the system, not the model

Two levers often beat “buy the biggest”:

· Better context assembly (structured sources over stale text; minimal, relevant snippets over long dumps).

· Routing by task difficulty (default to efficient models; escalate for complex or high-stakes cases).

Why it matters: You get higher reliability and lower cost (simultaneously) by improving how work is framed, not just who answers it.

5) Speed is part of trust

Even well-grounded answers won’t be used if they’re slow. Budget minimal context and choose authoritative sources to keep response times predictable. Think of latency like you think of page load in e-commerce: invisible until it becomes a problem. Adoption depends on flow. If it breaks flow, it won’t stick.

Where this leaves the “Is RAG dead?” debate

It’s the wrong question. The right question is: does your system assemble the right context, under the right rules, with visible receipts and can it keep doing that as models improve faster than you planned?

Hedden tells you how to answer “yes” today. Schrittwieser tells you why you’ll need to keep that “yes” resilient tomorrow.

Build for meaning and governance; plan for churn. That’s how digital transformation spend survives the hype cycle and benefits from the improvements still to come.

Hedden & Schrittwieser: Context Today, Churn Tomorrow