World 1-4: what holds the platforms up

Décio Dalke Jr. — Arquiteto de negócios. Sócio-gerente no ecossistema MitUP (mitup.pt).

World 1 of AI series — level 4 of 5: World 1-1 · World 1-2 · World 1-3 · World 1-4 · World 1-5

In World 1-3, Mario jumped from platform to platform high up in the air. There’s a question no player ever asks: what holds those platforms up? In the game, nothing. They’re drawn there. In a company, someone has to have built them.

A company’s AI architecture is the set of pieces that sit around the model: what it sees at any given moment, where it looks up what it doesn’t know, which tools it uses to act, how you measure whether it got it right and what stops it from doing something stupid. That — not the model you picked — is what decides whether AI works day to day.

This level is about those pieces. Anyone who has built an information system will recognise almost all of them. What’s new is that now the system talks back.

What does the AI see at any given moment?

World 1-2 showed that the model only knows what’s in front of it, in the context window. That gives us the first piece: deciding what goes into that window, and what stays out. Anthropic calls this context engineering and describes it as the set of strategies for choosing and keeping the best information in the window at each point in the work [1]. Notice the shift in focus: it stops being “how do I write the request” and becomes “what does it need to have at hand to answer well”.

In World 1-1 I told the story of my instructions, which kept growing until they became a single, generic block. The fix wasn’t to write them better. It was to split them by role, so that each task carried only what it needed. That’s context engineering.

How does the AI consult what doesn’t fit in the window?

Your company has more documents than any window can hold. The most common answer to that has an ugly acronym: RAG, retrieval-augmented generation. The name comes from a 2020 paper [2], and the idea is simple. Before answering, the system searches your documents for the relevant passages, puts those passages in the window and only then asks the model for an answer.

It’s the consultant with no memory from World 1-2, now with an archivist who pulls the right folder before he arrives.

The weak spot is just as simple: the answer is never better than the search. If the archivist hands over the old version of the contract, the consultant will answer with total confidence about the old version. That’s why the rule I mentioned in World 1-2, “one piece of information, one home”, isn’t a tidiness obsession. With three copies of the same document, RAG has three chances to grab the wrong one.

How does AI go from answering to acting?

With tools. A connector lets the model read your email, create a task, look something up in a spreadsheet, send a message. In November 2024 Anthropic launched an open standard for this, MCP (Model Context Protocol) [3]. It caught on: OpenAI announced support in March 2025 [4], and at the end of 2025 the standard was donated to a foundation under the Linux Foundation, backed by Google, Microsoft and AWS, among others [5].

For a manager, the sentence that matters is this one: every connector is a permission. Connecting AI to your email means handing it the keys to your email. And keys can be misused by third parties: in 2025, researchers showed that a malicious connector can hide instructions in its own description and lead the agent to leak data [6].

When is it worth having several agents work together?

The idea is seductive: one agent researches, another writes, another reviews. Sometimes it works very well. Anthropic itself showed that its multi-agent research system performed 90% better than a single agent in an internal evaluation, but it burns about 15 times more tokens than an ordinary chat [7]. On the other side, Cognition, which builds coding agents, published a piece titled, bluntly, “Don’t Build Multi-Agents”: decisions get scattered too widely, and context doesn’t pass whole from one agent to the next [8].

Both things are true at once. Anthropic’s advice for anyone starting out is to start simple and only add steps when simple can’t cope [9].

In my operation, multiple agents are worth it when each one has a role that can’t be mixed with another’s. The contract case, in World 1-1, is the example: a lawyer on our side, one for the other party and a judge. What went wrong was letting a single review, with no defined side, pick a side on its own.

How do you know it works?

This is the piece almost nobody builds, and it’s the one that brings projects down.

“I tried it a few times and it looked fine” isn’t a test. The name the industry uses for a real test is evaluation (eval): a set of cases with a known right answer, which you run again every time you change something [10]. Changed the instructions, swapped the model, added a document? Run the cases again and compare.

The best test cases aren’t made up. They come from mistakes that have already happened. In World 1-2 I said every rule in my method was born from something that went wrong. It’s the same move: yesterday’s mistake becomes the question the system has to get right tomorrow.

How do you stop AI from doing something stupid?

Asking doesn’t work. An instruction like “never do X” is a request, and the model may not follow it, or may be talked out of following it. OWASP, a benchmark in software security, lists “excessive agency” among the top risks of AI applications and recommends two things: give each tool only the minimum permission it needs, and run the authorisation check in the target system, instead of letting the model decide whether it’s allowed [11].

In other words: the limit has to sit outside the agent. The WhatsApp case, in World 1-3, is exactly this in practice. The agent stopped using the paid agent because its access was removed and the process was moved behind a gate that decides what it can do on its own.

Instructions, RAG or fine-tuning: where do you start?

It’s the question every manager ends up asking when AI doesn’t deliver what it should. The three options solve different problems:

If the problem is…	Start with…	Why
It didn’t understand what you wanted	Better instructions and context	It’s the cheapest and fastest to test
It doesn’t know the facts of your business	RAG (searching your documents)	The facts stay at the source, up to date and with their origin alongside
It knows, but doesn’t do it the right way, at volume	Fine-tuning (adjusting the model itself)	It teaches format, style and repeated behaviour

OpenAI’s documentation on model optimisation says that work on the instructions alone may be all you need [12]. And research points the same way for facts: in a study comparing the two techniques for teaching a model new knowledge, RAG consistently beat fine-tuning [13].

My reading, based on what I see in the market and not on measured data: for most companies, fine-tuning is rarely the first step. When AI gets facts wrong, fine-tuning tends to be the wrong answer to the right question. World 1-5 opens up that engine and shows why.

None of the pieces in this level show up on screen. That’s why they’re the first to be cut from the budget, and the first to be missed.

A platform nobody built won’t survive the first jump.

Quick questions

What is RAG, in one sentence? It’s having the system search your documents for the relevant passages before the model answers, so the answer rests on your data and not on the world’s average.

What is MCP? An open standard, launched by Anthropic in 2024, for connecting AI models to tools and data (email, files, systems). Today it’s adopted by several vendors.

Is it worth using multiple agents? When each agent has a role that can’t be mixed with the others’, yes. For everything else, start with just one: multiple agents cost a lot more and are harder to audit.

Does fine-tuning teach AI new facts? Poorly. It’s better for format, style and behaviour. For the facts of your business, the safer route is RAG.

References

1. Anthropic — Effective context engineering for AI agents, 29 September 2025. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents — accessed 2026-09-20.
2. Lewis, P. et al. — Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020. https://arxiv.org/abs/2005.11401 — accessed 2026-09-20.
3. Anthropic — Introducing the Model Context Protocol, 25 November 2024. https://www.anthropic.com/news/model-context-protocol — accessed 2026-09-20.
4. TechCrunch — OpenAI adopts rival Anthropic’s standard for connecting AI models to data, 26 March 2025. https://techcrunch.com/2025/03/26/openai-adopts-rival-anthropics-standard-for-connecting-ai-models-to-data/ — accessed 2026-09-20.
5. Model Context Protocol — MCP joins the Agentic AI Foundation, 9 December 2025. https://blog.modelcontextprotocol.io/posts/2025-12-09-mcp-joins-agentic-ai-foundation/ — accessed 2026-09-20.
6. Invariant Labs — MCP Security Notification: Tool Poisoning Attacks, 1 April 2025. https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks — accessed 2026-09-20.
7. Anthropic — How we built our multi-agent research system, 13 June 2025. https://www.anthropic.com/engineering/multi-agent-research-system — accessed 2026-09-20.
8. Cognition — Don’t Build Multi-Agents, 12 June 2025. https://cognition.com/blog/dont-build-multi-agents — accessed 2026-09-20.
9. Anthropic — Building effective agents, 19 December 2024. https://www.anthropic.com/engineering/building-effective-agents — accessed 2026-09-20.
10. Husain, H. — Your AI Product Needs Evals, 29 March 2024. https://hamel.dev/blog/posts/evals/index.html — accessed 2026-09-20.
11. OWASP — LLM06:2025 Excessive Agency (Top 10 for LLM Applications 2025). https://owasp.org/www-project-top-10-for-large-language-model-applications/2_0_vulns/LLM06_ExcessiveAgency.html — accessed 2026-09-20.
12. OpenAI — Model optimization (API documentation). https://developers.openai.com/api/docs/guides/model-optimization — accessed 2026-09-20.
13. Ovadia, O. et al. — Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs, 2023/2024. https://arxiv.org/abs/2312.05934 — accessed 2026-09-20.