Give agents a code sandbox instead of tool catalogs to eliminate context bloat
“Essentially, whatever you need, the coding agent can actually create the software for you dynamically.”
context-window
8 tracked signals on context-window.
Larger context windows cause agent hallucination; context engineering is the real fix.
“context engineering is helping us create a better architecture for these agent programs”
LangChain's Deep Agents harness uses offloading and summarization to combat agent context rot
“The more history, noise, and tool output a model carries, the worse its reasoning and accuracy tend to get. We call this context rot.”
LangChain's 'skills' pattern lazily loads agent context to prevent context window overload
“With skills, information is loaded in layers, not all at once, so context stays small until it's actually useful.”
LLM performance degrades after 25% context window usage, making 'infinite context' a dangerous myth
“if we consider that there's infinite context window, we just think that we can dump the whole list of documents to the context and expect it to magically work”
Use MCP servers for known, enterprise-auth, human-in-the-loop tools; use CLIs for model-familiar, light-surface tasks to save tokens.
“the MCP server for GitHub is about 45,000 tokens because it's giving every single API name, description, and schema to the model. But with the CLI, running the help command where it can see all of the tools, there's only about 1,500 tokens”
AWS Bedrock AgentCore enables unlimited document processing via Recursive Language Models
GitHub Copilot's 400k context window works best with minimal, targeted information rather than maximum context.
“Our goal isn't to give Copilot the most amount of context possible. It is to give it the smallest amount of information for it to accomplish its task.”