The Architecture of Information: Context Engineering as the New Frontier of AI Productivity

At Money Forward Vietnam (MFV), being an "AI-first" company means much more than just typing better prompts into a chat interface. We are building AI at the architectural level. While the industry is still fixated on "Prompt Engineering", our engineering teams have already transitioned to a more rigorous, scalable discipline: Context Engineering. In this deep dive, MFV AI Engineer Henry Vo breaks down the architecture of information, the "pathologies" of AI context windows, and how MFV is fundamentally changing the software development lifecycle.

Part 1: The Vision - A Cognitive Shift

The evolution of artificial intelligence from simple chat interfaces to sophisticated, autonomous digital workers has precipitated a fundamental shift in software engineering priorities. For several years, the industry remained fixated on prompt engineering—the specialized craft of phrasing instructions to elicit specific responses from large language models. However, as organizations like Money Forward move toward an "AI-native" development paradigm, a more rigorous and scalable discipline has emerged: context engineering ¹.

This discipline transcends the linguistic nuances of prompting, focusing instead on the technical architecture of the information environments in which AI models operate. It treats the model’s context window not as a static text box, but as a dynamic and scarce computational resource analogous to random-access memory (RAM) in traditional computing architectures ³. The transition from "how we ask" to "what information we provide" is the defining characteristic of the current era of AI agents¹.

Part 2: The Reality - Pathologies of the Context Window

The limitations of traditional prompt engineering become apparent when moving beyond simple, single-turn interactions toward complex, multi-step agentic workflows. Engineers must identify and mitigate four distinct "pathologies" that cause AI agents to behave unreliably in production environments:

• Context Poisoning: Poisoning occurs when an error or a hallucination enters the context and is repeatedly referenced by the model as a factual baseline for future reasoning.
• Context Distraction: Distraction happens when a context grows so long that the model over-focuses on recent history or redundant information, neglecting its foundational training and system-level instructions.
• Context Confusion: Confusion arises when superfluous or irrelevant content—such as too many tool definitions or irrelevant code snippets—is placed in the prompt.
• Context Clash: Clash occurs when parts of the context disagree with one another. This is frequent in multi-turn interactions where early, incorrect attempts by the model remain in the history and conflict with later, correct information.

Part 3: The Capability - The Four Pillars of Context Engineering

To solve these pathologies, developers must implement a comprehensive framework for managing the information lifecycle. This framework consists of four core pillars: Write, Select, Compress, and Isolate ³.

• Pillar 1: Write (State Persistence): Writing context involves saving critical information externally to bypass the physical limits of the context window ³. Agents use scratchpads to record plans, intermediate thoughts, or state data during a task ³.
• Pillar 2: Select (Just-in-Time Retrieval): Selection is the process of pulling only the most relevant tokens into the context window for a specific inference step ¹. Rather than loading an entire codebase, the agent uses tools to "discover" information dynamically¹⁵.
• Pillar 3: Compress (Token Efficiency): Compression focuses on reducing the size of information in the context window while preserving its semantic value³.
• Pillar 4: Isolate (Cognitive De-loading): Isolation involves dividing information across multiple agents or environments to prevent interference and maintain focus³.

Part 4: The Impact - AI-Native Evolution at Money Forward

Money Forward has implemented context engineering at scale. The deployment of tools like Cursor and Claude Code has transformed how different roles within Money Forward interact with the codebase and with each other. This transformation is driven by the principles of context engineering: providing the right data to the right person (or agent) at the right time⁸.

The productivity impacts are already clear across our teams:

• QA Engineers: Use MCP to feed agents Jira tickets and Notion docs for automated test generation, leading to a 70% reduction in time spent on test case generation ⁸.
• Engineers: Leverage Cursor's context retrieval for massive projects like Rails optimization and legacy migrations, saving individual developers an estimated 15–20 hours per week ²³.

Part 5: Best Practices: Designing Tools as Part of the Cognitive Environment

Designing Tools as Part of the Cognitive Environment

In the world of context engineering, tool definitions are not just API endpoints; they are a critical part of the model’s prompt.3 Effective tool design is essential for maintaining the "Goldilocks zone" of the system prompt—specific enough to guide behavior but flexible enough to allow for model reasoning.¹⁵

Best Practices for Agentic Tool Design

Prioritize Token Efficiency: Tools should return meaningful, semantically rich data rather than technical boilerplate³.Returning 1,000 rows of raw JSON is a failure of context engineering; returning a summary of the top 5 relevant items is a success³.

Use Descriptive Parameters: Instead of naming a parameter id, name it user_id_to_fetch³. Clear naming conventions act as "micro-prompts" that guide the agent’s reasoning about how to use the tool correctly³.

Return Actionable Error Messages: When a tool call fails, the response should guide the agent toward a solution³. For example: "Search yielded too many results. Please narrow your search by adding a 'category' filter"³.

Avoid Technical IDs: Agents think in concepts, not hashes. Whenever possible, tools should return names, titles, or descriptions alongside technical IDs to help the agent maintain its "mental model" of the task³.

Designing Tools as Part of the Cognitive Environment

In the world of context engineering, tool definitions are not just API endpoints; they are a critical part of the model’s prompt³. Effective tool design is essential for maintaining the "Goldilocks zone" of the system prompt—specific enough to guide behavior but flexible enough to allow for model reasoning¹⁵.

Best Practices for Agentic Tool Design

1. Prioritize Token Efficiency: Tools should return meaningful, semantically rich data rather than technical boilerplate.3 Returning 1,000 rows of raw JSON is a failure of context engineering; returning a summary of the top 5 relevant items is a success³.
2. Use Descriptive Parameters: Instead of naming a parameter id, name it user_id_to_fetch³. Clear naming conventions act as "micro-prompts" that guide the agent’s reasoning about how to use the tool correctly³.
3. Return Actionable Error Messages: When a tool call fails, the response should guide the agent toward a solution³. For example: "Search yielded too many results. Please narrow your search by adding a 'category' filter"³.
4. Avoid Technical IDs: Agents think in concepts, not hashes. Whenever possible, tools should return names, titles, or descriptions alongside technical IDs to help the agent maintain its "mental model" of the task³.

The future of software is not just about writing better code; it is about building better information architectures ⁶. As agents evolve into sophisticated digital colleagues, the ability to engineer their context will be the most critical skill for the modern developer¹.

Are you ready to stop prompting and start engineering? If you want to build at the architectural level and solve high-scale AI challenges, MFV is your ultimate playground. Join us in building the future of FinTech.

Works cited and read more

1. Why AI Teams Are Moving From Prompt Engineering to Context Engineering - Neo4j, accessed March 26, 2026, https://neo4j.com/blog/agentic-ai/context-engineering-vs-prompt-engineering/
2. Context Engineering 101
3. Context Engineering - LangChain Blog, accessed March 26, 2026, https://blog.langchain.com/context-engineering-for-agents/
4. Context Engineering: The Real Advantage in Generative AI - Blog de Bismart, accessed March 26, 2026, https://blog.bismart.com/en/context-engineering-vs-prompt-engineering-generative-ai
5. The Evolution of Prompt Engineering to Context Design in 2026 - SDG Group, accessed March 26, 2026, https://www.sdggroup.com/en/insights/blog/the-evolution-of-prompt-engineering-to-context-design-in-2026?hsLang=en
6. Money Forward - HashiCorp, accessed March 26, 2026, https://www.hashicorp.com/en/case-studies/money-forward
7. Money Forward Charges Into 2026 with Strategic Acquisitions and Aggressive AI Deployment - Japan FinTech Observer, accessed March 26, 2026, https://www.fintechobserver.com/money-forward-charges-into-2026-with-strategic-acquisitions-and-aggressive-ai-deployment/
8. humanlayer/12-factor-agents: What are the principles we ... - GitHub, accessed March 26, 2026, https://github.com/humanlayer/12-factor-agents
9. The 12-Factor Agent: A Practical Framework for Building Production AI Systems, accessed March 26, 2026, https://dev.to/bredmond1019/the-12-factor-agent-a-practical-framework-for-building-production-ai-systems-3oo8
10. Operationalizing AI Agents: Lessons from 2025 | Xpress AI, accessed March 26, 2026, https://xpress.ai/blog/2025-agent-lessons
11. Context engineering in agents - Docs by LangChain, accessed March 26, 2026, https://docs.langchain.com/oss/python/langchain/context-engineering
12. From Chaos to Clarity: My Journey into the World of 12-Factor Agents | by HIYA CHATTERJEE | Python in Plain English, accessed March 26, 2026, https://python.plainenglish.io/from-chaos-to-clarity-my-journey-into-the-world-of-12-factor-agents-cb6eca360a42
13. Context Engineering : Critical Shift from Prompting to Engineering - Futran Solutions, accessed March 26, 2026, https://futransolutions.com/blog/context-engineering-the-critical-shift-from-prompting-to-engineering/
14. Effective context engineering for AI agents - Anthropic, accessed March 26, 2026, https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
15. Effective context engineering for AI agents \ Anthropic, accessed March 26, 2026, https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents/
16. 12 Factor Agents: Principles for AI That Actually Work - Emergent Minds | paddo.dev, accessed March 26, 2026, https://paddo.dev/blog/12-factor-agents/
17. Context engineering in Deep Agents - Docs by LangChain, accessed March 26, 2026, https://docs.langchain.com/oss/python/deepagents/context-engineering
18. Tech.IT Forward #5 - Q&A Recap, accessed March 26, 2026, https://careers.moneyforward.vn/blog/techit-forward-5-qa-recap
19. Money Forward's English Transformation: A Blueprint for Engineering Teams in Japan in 2025, accessed March 26, 2026, https://japan-dev.com/blog/money-forwards-english-transformation-a-blueprint-for-engineering-teams-in-japan-in-2025
20. AI Initiatives｜Money Forward, Inc., accessed March 26, 2026, https://corp.moneyforward.com/en/service/ai/
21. Money Forward brings Cursor's coding agents to product, design, and QA, accessed March 26, 2026, https://cursor.com/blog/money-forward
22. Developer's Guide to AI Agent Protocols, accessed March 26, 2026, https://developers.googleblog.com/developers-guide-to-ai-agent-protocols/
23. Use Agents or Be Left Behind? A Personal Guide to Automating Your Own Work, accessed March 26, 2026, https://timdettmers.com/2026/01/13/use-agents-or-be-left-behind/