# Agent Context Engineering with File Systems

***

## What Is Agent Context Engineering?

**Agent context engineering** is the discipline of deliberately managing what information an AI agent sees at each step of its reasoning and action loop.

It goes beyond prompt engineering and includes:

* External memory (files, databases, vector stores)
* Retrieval and selection strategies
* Compression and summarization
* Context isolation and lifecycle management

The goal is to ensure the agent always has **the right context, at the right time, in the right format**.

***

## Why Context Engineering Matters (5 Reasons)

1. **Limited Token Context**
   * LLMs have bounded context windows.
   * Poor context selection leads to truncation or confusion.
2. **Performance & Cost**
   * Smaller, relevant context = faster inference + lower cost.
   * Avoids dumping large tool outputs directly into prompts.
3. **Avoiding Context Rot**
   * Too much irrelevant or outdated information degrades reasoning quality.
   * Common failure mode observed in long-running agents.
4. **Persistent Memory**
   * File systems allow cross-step and cross-session memory.
   * Enables agents to “remember” without bloating the prompt.
5. **Traceability & Debugging**
   * Externalized context (files, artifacts) is inspectable and reproducible.
   * Critical for debugging agent failures and auditing behavior.

***

## Key Sources & Insights

### Manus — Context Engineering Lessons

**Source:**

* [Manus Blog - Context Engineering for AI Agents](https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus)
* [how-manus-uses-e2b-to-provide-agents-with-virtual-computers](https://e2b.dev/blog/how-manus-uses-e2b-to-provide-agents-with-virtual-computers)

**Key Ideas:**

* Context engineering is more impactful than model tuning.
* Treat the **filesystem as unbounded memory**.
* Write large observations to files; load selectively.
* Reuse stable prefixes and cache-friendly context.

**Filesystem Role:**

* Store long tool outputs, documents, and intermediate results.
* Retrieve only what is needed for the next step.

![Manus Context Engineering - Filesystem as Unbounded Memory](https://388701358-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F-LmDY11BLFD0Iupj9U9t%2Fuploads%2Fgit-blob-bc0409f4dddad0afaae8647f7e7bb1f570427a49%2Fmanus-context-engineering.png?alt=media)

![Manus Summary - Context Engineering Overview](https://388701358-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F-LmDY11BLFD0Iupj9U9t%2Fuploads%2Fgit-blob-2e0289c35e9ec39105fa50a84de2dc80011b2fb5%2Fmanus-summary-context-engineering.png?alt=media)

***

### Anthropic — Effective Context Engineering

**Source:** [Anthropic - Effective Context Engineering for AI Agents](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents)

**Key Ideas:**

* Context is a finite, valuable resource.
* Focus on holistic context optimization, not just prompts.
* Carefully curate instructions, tool outputs, memory, and history.

![Anthropic - Effective Context Engineering](https://388701358-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F-LmDY11BLFD0Iupj9U9t%2Fuploads%2Fgit-blob-57855cc7db56a899e5921de2973daf02be279f0c%2Fanthropic-context-engineering.png?alt=media)

***

### LangChain — Context Engineering for Agents

**Source:** [LangChain Blog - Context Engineering for Agents](https://blog.langchain.com/context-engineering-for-agents/)

**Four Core Strategies:**

* **Write:** Save context externally
* **Select:** Retrieve relevant information
* **Compress:** Summarize or prune
* **Isolate:** Separate concerns and states

![Agent Context Engineering Legend - Four Core Strategies](https://388701358-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F-LmDY11BLFD0Iupj9U9t%2Fuploads%2Fgit-blob-e5201a11daa901f2d7ed656afe2523278b1ffaca%2Fagent-context-engineering-legend.png?alt=media)

***

### LangChain — Filesystems for Agents

**Source:** [LangChain Blog - How Agents Can Use Filesystems](https://blog.langchain.com/how-agents-can-use-filesystems-for-context-engineering/)

**Key Ideas:**

* Filesystems act as flexible external memory.
* Agents write structured artifacts (search results, plans, logs).
* Files can be indexed, summarized, or selectively reloaded.

![LangChain - Filesystems for Context Engineering](https://388701358-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F-LmDY11BLFD0Iupj9U9t%2Fuploads%2Fgit-blob-f6d087e38cf45203bfae0a13e7ee1c7530d15149%2Flangchain-context-engineering-file-system.png?alt=media)

***

### Cursor — Dynamic Context Discovery

**Source:** [Cursor Blog - Dynamic Context Discovery](https://cursor.com/blog/dynamic-context-discovery)

**Key Ideas:**

* Do not preload all context.
* Discover and load context **just in time**.
* Convert tool outputs into file artifacts.
* Pull relevant slices only when needed.

![Dynamic Context Discovery - Static vs Dynamic Context](https://388701358-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F-LmDY11BLFD0Iupj9U9t%2Fuploads%2Fgit-blob-521afc58f874b38e0ffe92dfca309e395a866201%2Fdynamic-context-discovery.jpg?alt=media)

### Vercel - How to build agents with filesystems and bash

**Source:** [how-to-build-agents-with-filesystems-and-bash](https://vercel.com/blog/how-to-build-agents-with-filesystems-and-bash)

***

## Academic Research

### “Everything is Context: Agentic File System Abstraction for Context Engineering”

**Paper:** [arXiv:2512.05470 - Everything is Context](https://arxiv.org/abs/2512.05470)

**Contributions:**

* Proposes a Unix-like filesystem abstraction for agent context.
* All context artifacts become governed files.
* Enables:
  * Metadata-aware access
  * Traceability
  * Reproducibility
  * Structured context pipelines

![Everything is Context - Agentic File System Abstraction](https://388701358-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F-LmDY11BLFD0Iupj9U9t%2Fuploads%2Fgit-blob-e1eb74e66baa2ece7a8398fbc378102483112ef3%2Feverything-is-context-file-system.png?alt=media)

***

### “Agent READMEs: An Empirical Study”

**Paper:** [arXiv:2511.12884 - Agent READMEs](https://arxiv.org/abs/2511.12884)

**Insights:**

* Context files behave like evolving configuration and instruction artifacts.
* They strongly influence agent behavior over time.

![Agent Coding Context - Filesystem-Based Pattern](https://388701358-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F-LmDY11BLFD0Iupj9U9t%2Fuploads%2Fgit-blob-310c5041fbba8a7d9ae0c2a35c7520bef460a974%2Fagent-coding-context-file.png?alt=media)

***

## Timeline: Evolution of Context Engineering

### 2023–2024: Prompt Engineering Era

* Static prompts dominate.
* Minimal external memory usage.

### 2025 Q1–Q2: Context Engineering Defined

* Anthropic & LangChain formalize the discipline.
* Token budget awareness becomes mainstream.

### 2025 Q2–Q3: Filesystems & Practical Lessons

* Manus demonstrates filesystem-based memory.
* Practitioners identify context rot issues.

### 2025 Q4: Dynamic Context Discovery

* Cursor introduces just-in-time context loading.
* File-based artifact discovery gains popularity.

### Dec 2025: Academic Formalization

* “Everything is Context” paper introduces agentic file systems.

### 2026: Adaptive Memory Agents

* Layered memory
* Self-curating context
* Learning what context to retrieve

***

## Core Pattern: Filesystem-Based Context Engineering

```
Agent Step
  ↓
Tool Call / Observation
  ↓
Write to File (Artifact)
  ↓
Index / Tag / Summarize
  ↓
Selective Retrieval
  ↓
Inject into Context Window
```


---

# Agent Instructions: Querying This Documentation

If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter:

```
GET https://snowan.gitbook.io/study-notes/ai/ai-resources-1/context-engineering/context-engineering-with-file-system.md?ask=<question>
```

The question should be specific, self-contained, and written in natural language.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.