Context Engineering: From Prompts to Corporate Multi-Agent Architecture

This paper proposes "Context Engineering" as a foundational discipline that, alongside Intent and Specification Engineering, forms a maturity model for scaling autonomous multi-agent systems by shifting focus from individual prompts to the comprehensive management of an agent's informational environment, goals, and policy constraints.

The Gist

Imagine you are the captain of a massive, high-tech spaceship. In the past, you had to manually steer the ship, push every button, and shout every command to the engine room. That was the era of Prompt Engineering. You gave a specific order ("Go to Mars"), and the ship did exactly that, one step at a time.

But now, the ship has evolved. It has a crew of autonomous robots (Agents) that can plan routes, fix engines, and talk to each other without you shouting every single instruction. The problem? If you just keep shouting orders like you used to, the ship will crash. The robots need more than just a command; they need a system.

This paper argues that to manage these robot crews, we need to upgrade our thinking from "giving orders" to "building the environment" in which they live. The author, Vera Vishnyakova, proposes a Four-Level Pyramid of skills needed to run these AI systems successfully.

Here is the breakdown of the four levels, using simple analogies:

Level 1: The Art of Asking (Prompt Engineering)

The Analogy: The Magic Wand.
In the beginning, you just waved a magic wand and said, "Make me a sandwich." If you said it clearly, you got a sandwich. If you said it poorly, you got a mess. This is Prompt Engineering.

• What it is: Crafting the perfect question or instruction for an AI.
• The Limit: It works great for one-off tasks. But if you tell a robot to "run a whole company," it can't do that with just one magic phrase. It needs a system to handle 50 steps, not just one.

Level 2: The Operating System (Context Engineering)

The Analogy: The Chef's Kitchen.
Imagine you hire a brilliant chef (the AI). If you just tell them "Make a cake," but you don't give them the ingredients, the recipe, the oven temperature, or the dietary restrictions of the guests, they will fail.
Context Engineering is about building the kitchen. It's not about the order you give; it's about what the chef sees and remembers while cooking.

• Relevance: Don't give the chef a whole library of books; just give them the recipe for a cake. (Too much info confuses them).
• Isolation: If you have two chefs, Chef A shouldn't see Chef B's dirty dishes. They need their own clean workspace.
• Economy: Don't waste money buying 100 pounds of flour if you only need 2 pounds.
• The Problem Solved: Without this, the AI gets "lost in the middle," forgets what it was doing, or mixes up data from different tasks.

Level 3: The Company's Soul (Intent Engineering)

The Analogy: The Compass and the Map.
Imagine you give the robot chef a perfect kitchen with perfect ingredients (Level 2). But you forgot to tell them what kind of cake the customer actually wants.
The robot might make a delicious, expensive, 10-layer chocolate cake. But the customer is on a diet and just wanted a small muffin. The robot did its job perfectly, but it failed the goal.
Intent Engineering is about encoding the company's values and goals. It answers: "What matters more? Speed or quality? Saving money or making the customer happy?"

• The Klarna Warning: The paper mentions a real company (Klarna) that saved millions by using AI to answer customer calls. But the AI was too blunt and rude because no one told it, "Be polite and keep customers happy." It optimized for "speed" and "cost" and accidentally ruined the brand.
• The Fix: You have to program the AI's "moral compass" so it knows why it is doing the task, not just how.

Level 4: The Constitution (Specification Engineering)

The Analogy: The Rulebook for a City.
Now imagine you don't just have one robot chef; you have 10,000 robots working in a giant factory. If every robot makes up its own rules, chaos ensues. One robot thinks "fast" is good; another thinks "cheap" is good. They start fighting.
Specification Engineering is writing the "Constitution" or the "Rulebook" for the entire city of robots. It turns all the messy, unwritten company rules (like "we value safety") into a strict, machine-readable code that every robot must follow.

• Why it matters: You can't manage 10,000 robots with a verbal agreement. You need a digital law book that ensures they all work together coherently.

---

The Big Picture: The Pyramid

The paper says these four levels are like a pyramid. You can't build the top without the bottom.

1. Base: You must know how to ask good questions (Prompt Engineering).
2. Middle: You must build a good environment for them to work in (Context Engineering).
3. Upper: You must give them a clear purpose and values (Intent Engineering).
4. Top: You must write a rulebook for the whole system to scale up (Specification Engineering).

The "Dark Factory" Trap

The author warns us about a dangerous trend. Because it's now so easy to create AI agents (like building a Lego set), companies are rushing to build them without thinking about Levels 2, 3, and 4.
They build a robot that works technically but has no values, no clear rules, and a messy memory. This creates a "Dark Factory": a place where robots are working autonomously, but no human really understands what they are doing or why. They might be saving money today, but tomorrow they could destroy the company's reputation or break the law.

The Takeaway

To succeed with AI in the future, we can't just be "prompt whisperers" (people who are good at talking to AI). We need to become Architects.

• We need to design the environment (Context).
• We need to define the purpose (Intent).
• We need to write the laws (Specifications).

If you control the context, you control the behavior. If you control the intent, you control the strategy. If you control the specifications, you control the scale. The paper concludes that as AI gets smarter, our job isn't to do the work for the machine, but to design the world in which the machine works.

Technical Summary

Based on the paper "Context Engineering: From Prompts to Corporate Multi-Agent Architecture" by Vera V. Vishnyakova (March 2026), here is a detailed technical summary.

1. Problem Statement

The paper identifies a critical architectural gap in the evolution of AI from stateless chatbots to autonomous multi-step agents. While Prompt Engineering (PE) is effective for single-turn "human-to-model" interactions, it is insufficient for Multi-Agent Systems (MAS) and autonomous agents that operate over long horizons (20–50+ steps).

Key problems identified include:

• Context Degradation: In long-running agent sessions, models suffer from "lost-in-the-middle" phenomena, context window crowding, and contamination from irrelevant tool outputs.
• Lack of Isolation: In MAS, data leakage between sub-agents leads to decision-making based on unauthorized or irrelevant information.
• Economic Unsustainability: Without compression and caching, token costs and latency grow super-linearly (super-linearly to quadratically) with the number of agent steps.
• The Intent Deficit: Even with perfect context, agents may optimize for the wrong metrics (e.g., cost per token) if corporate goals, values, and trade-off hierarchies are not explicitly encoded.
• Governance Gap: Enterprises are deploying agents rapidly (75% plan within two years) but lack the governance structures to manage what agents "see" (context) and "want" (intent), leading to reputational and strategic risks (exemplified by the Klarna case).

2. Methodology

The paper employs a synthesis of industry practice, vendor architecture analysis, and academic research to propose a new engineering framework.

• Data Sources: Analysis of vendor architectures (Google ADK, Anthropic, LangChain), academic frameworks (ACE, Tomašev et al. 2026), and enterprise survey data (Deloitte 2026, KPMG 2026).
• Case Studies: Examination of real-world deployments, including a multi-agent compliance system built by the author, the Klarna AI customer service failure, and the TELUS scale challenge.
• Conceptual Framework: The author introduces a Cumulative Four-Level Pyramid Maturity Model, arguing that higher levels of engineering do not replace lower ones but subsume them as necessary infrastructure.
• Taxonomy Development: The paper defines specific quality criteria for context and degradation modes (Context Rot) based on observed failure patterns in production environments.

3. Key Contributions

A. The Four-Level Maturity Pyramid

The paper proposes a hierarchical model of Agent Engineering consisting of four cumulative disciplines:

1. Prompt Engineering (PE): The foundation. Crafting queries for stateless, single-turn interactions. Still necessary but insufficient for autonomous agents.
2. Context Engineering (CE): The core focus. The design and management of the agent's "operating system"—the informational environment. It manages memory, policies, tool outputs, and visibility boundaries.
3. Intent Engineering (IE): Encoding organizational goals, values, and trade-off hierarchies. It answers "what the agent wants" and ensures strategic alignment, preventing reward hacking.
4. Specification Engineering (SE): The highest level. Creating a machine-readable corpus of corporate policies, standards, and agreements (a "Constitution for Agents") to enable coherent scaling of multi-agent systems.

B. Five Production-Grade Context Quality Criteria

The paper defines five criteria that production-grade agent context must satisfy:

1. Relevance: Only data necessary for the current step is fed to the agent (Just-In-Time knowledge logistics).
2. Sufficiency: The context must contain all data required to make a decision without hallucination.
3. Isolation: Sub-agents must see only their specific slice of context to prevent data leakage and privilege escalation (implemented via Delegation Capability Tokens).
4. Economy: Minimizing token usage and context reassembly through compression, caching, and selective loading to ensure unit economics viability.
5. Provenance: Every context element must be traceable to its source for auditing, debugging, and regulatory compliance.

C. Context Engineering as an Operating System

The paper reframes context not as passive input data but as an active execution environment (Operating System). It manages memory (retention/eviction), allocates resources, isolates processes, and provides interfaces to external systems. This shifts the discipline from "craft" (wording) to "systems engineering."

D. Diagnostic Taxonomy: Context Rot

The paper introduces a taxonomy for context degradation (based on Breunig, 2025):

• Poisoning: False facts entering context and propagating.
• Distraction: Over-reliance on accumulated history rather than current synthesis.
• Confusion: Irrelevant information degrading response quality.
• Clash: Contradictory data accumulating over time.

E. Memory Architecture Framework

The paper categorizes agent memory into four types, distinguishing between:

• Working Memory: Current context window (expensive, volatile).
• Episodic Memory: Logs of past interactions (compressed, long-term).
• Semantic Memory: Structured knowledge bases (RAG).
• Procedural Memory: Model weights and fine-tuned capabilities.
  It highlights the shift from user-managed memory to infrastructure-managed memory, where cloud providers dictate isolation and billing models.

4. Results and Findings

• The Klarna Case Analysis: The paper attributes Klarna's service quality decline to a dual deficit: a lack of Context Engineering (inadequate access to customer history/brand tone) and a lack of Intent Engineering (optimizing for cost rather than customer loyalty).
• Economic Impact: Sound Context Engineering can reduce inference costs by 5–10× through compression and caching, making SaaS agent products economically viable.
• Governance Reality: 84% of companies have not redesigned roles for AI, and only 21% have mature governance. The paper argues this is a "Specification Engineering" gap; companies cannot govern what they have not formalized into machine-readable specifications.
• Hybrid Architectures: The paper validates the trend of Cloud-Edge Hybrid MAS, where a cloud LLM acts as an orchestrator for edge-deployed Small Language Models (SLMs), requiring federated memory architectures to handle data sovereignty and latency constraints.

5. Significance and Implications

• Paradigm Shift: The paper marks the transition from "Prompt Engineering" (an art/craft) to "Context/Intent/Specification Engineering" (disciplines of systems engineering). It argues that the "romantic ceiling" of prompt engineering has been reached; further progress requires engineering rigor.
• Strategic Control: The paper concludes with a powerful managerial axiom:
  • Control of Context = Control of Behavior.
  • Control of Intent = Control of Strategy.
  • Control of Specifications = Control of Scale.
• Future-Proofing: The framework is presented as architecture-agnostic. Even if LLMs are replaced by world models or other AI architectures, the need for managing the informational environment (Context), strategic goals (Intent), and formal rules (Specifications) will intensify.
• Corporate Maturity: The Pyramid serves as a diagnostic tool for organizations. Stopping at Level 1 or 2 leads to "Dark Factory" scenarios (autonomous but uncontrolled agents), while reaching Level 4 enables a "machine-readable operating system" for the corporation.

In summary, the paper provides the theoretical and practical foundation for scaling AI from experimental chatbots to reliable, governed, and economically viable corporate multi-agent systems.