rohit.vision

Context Engineering & Filesystems

WIP nlp-llms agents context-engineering filesystem llm llm-agents memory nlp scratchpad 2 min read

How deep agents use filesystems as scratchpads for dynamic context engineering

Deep Agents (LangChain)

“Deep Agents” also refers specifically to the open-source DeepAgents repository by LangChain.

  • Architecture: It acts as an opinionated “Agent Harness” built on top of LangGraph. Instead of manually wiring up prompts and tools, you get a pre-configured agent capable of handling long-running, multi-step workflows.
  • Key Capabilities: It comes pre-bundled with Planning tools (write_todos), full Filesystem access (ls, grep, edit_file), and Sub-Agent delegation out of the box.

    Overview

Agents frequently fail not because the underlying LLM is bad, but because they retrieve either too much or too little context. As detailed in LangChain’s blog on Context Engineering, relying exclusively on the chat history buffer or pure semantic search leads to catastrophic context management failures.

  • Context Bloat: A web search tool might return 10k tokens. Putting all 10k tokens directly into the conversation history inflates API costs, degrades reasoning, and risks hitting token limits.
  • Niche Information: Relying purely on semantic vector search often fails to find highly specific, non-semantic facts (like exact code snippets or precise API arguments).

Filesystems as Agent Context

Modern “Deep Agents” (e.g., Manus, Claude Code) bypass the conversation history bottleneck by treating a Filesystem as a dynamic context scratchpad.

1. Offloading Tool Results

Instead of injecting a massive 10k token web search directly into the chat history, the agent writes the tool result to a file on disk. The agent can then use grep or standard file reading tools to selectively pull out only the exact 100-200 tokens it actually needs to answer the user’s prompt.

2. Replacing Semantic Search with Glob/Grep

For complex repositories, filesystems allow agents to use deterministic terminal commands (ls, glob, grep) to hunt down exact file paths or variable names rather than hoping a Vector Database clusters the code correctly.

3. Self-Improving System Prompts

If a user corrects an agent (“always use PyTest for this project”), the agent can write that preference into a instructions.txt file within its own filesystem. Future agent runs read that file upon initialization, effectively allowing the agent to dynamically update its own instructions without developers hard-coding new system prompts.