How Cursor’s Dynamic Context Cuts Agent Token Use by 47%

Cursor’s agent now uses dynamic context for all models, intelligently filling the context window and reducing total token usage by 46.9% while maintaining response quality.

Files Used for Dynamic Context Discovery

Dynamic context discovery is far more token‑efficient because only necessary data is brought into the context window, and it reduces confusing or contradictory information, improving the agent’s replies.

1. Convert Long Tool Responses to Files

Tool calls can return massive JSON payloads that bloat the context window. For Cursor’s built‑in tools (e.g., file editing, code‑base search) we can streamline responses, but third‑party tools like shell commands or MCP calls lack such optimization. Instead of truncating output, Cursor writes the output to a file and gives the agent the ability to read it. The agent uses tail to inspect the end of the file and fetch more content if needed, reducing unnecessary summarization when the context limit is approached.

2. Reference Conversation History During Summarization

When the model’s context window fills up, Cursor triggers a summarization step that provides the agent with a fresh window containing a summary of its work so far. Because this compression is lossy, the agent may forget critical details. Cursor therefore supplies the full conversation history as a file, allowing the agent to retrieve missing information on demand.

3. Support Agent Skills Open Standard

Cursor supports Agent Skills, an open standard for extending coding agents with specialized capabilities. Skills are defined in files that describe how to perform a specific task and can be included as static context in the system prompt. During execution, the agent can dynamically discover relevant skills using tools such as grep or Cursor’s semantic search, automatically pulling in the needed skill definitions.

4. Efficiently Load Only Required MCP Tools

MCP (Multi‑Channel Proxy) provides access to OAuth‑protected resources like production logs, design files, or internal documentation. Many MCP servers expose numerous tools with long descriptions, inflating the context window even when most tools are unused. Cursor synchronizes tool descriptions into a folder and gives the agent a small static context (just the tool names). The agent fetches detailed descriptions only when required. In an A/B test, this strategy reduced total token consumption by 46.9% (statistically significant, though variance depends on the number of installed MCP servers).

This file‑based approach also lets the agent convey the status of MCP tools, prompting users to re‑authenticate when a server’s credentials expire.

5. Treat All Integrated Terminal Sessions as Files

Previously, users had to copy‑paste terminal output into the agent’s input. Cursor now automatically syncs integrated terminal output to the local file system, allowing the agent to reference commands like “Why did my command fail?” and to grep only the relevant portions of long logs, which is especially useful for long‑running server processes.

This mirrors the behavior of CLI‑based coding agents, but the context is discovered dynamically rather than injected statically.

Simple Abstraction

It remains unclear whether files will become the ultimate interface for LLM‑based tools, but given the rapid evolution of coding agents, files are a simple and powerful primitive that avoids premature abstraction layers.