Improved token efficiency for longer agent runs · Cursor

Cursor Blog · Published Sep 23, 2026

The authors describe token-efficiency work in Cursor’s agent harness. They report a 7% reduction in user token costs without reducing agent quality, a roughly 66% reduction in the system prompt, a 46.9% reduction in total tokens for sessions calling an MCP tool after dynamic tool loading, and a 20% reduction in cold cache misses after cache-related changes. Read 4 viewpoints with supporting evidence and source links.

Jediah Katz, Connor O’Keefe & Calvin Yee

Read key viewpoints Open Publisher

Understand this piece

4 key points

Synthesis

7% token cost reduction without quality loss
Improvements to Cursor's agent harness across multiple layers reduced user token costs by 7% while maintaining agent quality.
Supporting evidence 1
Original excerpt
Changes across each of these layers reduced token costs for users by 7% without reducing agent quality.
Jediah Katz, Connor O’Keefe & Calvin Yee · Paragraph 7
Improved token efficiency for longer agent runs · Cursor · Sep 23, 2026
Context
Over the past few months we've responded to this shift by improving the efficiency of Cursor's agent harness. The harness gives us direct control over how each request is assembled, how context is reused, and when work is divided across agents.
Read in source context →
Continue exploring
token efficiency optimization →
66% system prompt reduction enabled by model capability gains
As models improved, explicit prohibitions and prescriptive instructions became unnecessary; Cursor trimmed approximately 66% of its system prompt by relying on model compliance with concise tool behavior definitions.
Supporting evidence 1
Original excerpt
This was true across model families, allowing us to trim roughly 66% of our system prompt.
Jediah Katz, Connor O’Keefe & Calvin Yee · Paragraph 11
Improved token efficiency for longer agent runs · Cursor · Sep 23, 2026
Context
As models improved, much of that direction became unnecessary. Instead of long lists of "DO NOT do this," "You must," or "Important" instructions, we could simply define how a tool behaves and models would generally comply.
Read in source context →
Continue exploring
system prompt design →
46.9% token reduction from dynamic MCP tool loading
Moving MCP tools into dynamic context—loading them only when invoked—reduced total tokens by 46.9% across sessions that used an MCP tool.
Supporting evidence 1
Original excerpt
This reduced total tokens by 46.9% across sessions that called an MCP tool.
Jediah Katz, Connor O’Keefe & Calvin Yee · Paragraph 15
Improved token efficiency for longer agent runs · Cursor · Sep 23, 2026
Context
That created an opportunity to improve efficiency by keeping tools available without including their full definitions in every request. We'd solved a similar problem earlier this year when we moved MCP tools into dynamic context, loading them only when needed.
Read in source context →
Continue exploring
dynamic context loading →
Cache-related changes reduced cold cache misses by 20%
Cache-related changes reduced cold cache misses by 20%. The authors reserved tools and system instructions for content that rarely changes, then moved more variable setup past the cache boundaries into a “phantom user message.”
Supporting evidence 1
Original excerpt
These changes reduced the rate of cold cache misses by 20%.
Jediah Katz, Connor O’Keefe & Calvin Yee · Paragraph 35
Improved token efficiency for longer agent runs · Cursor · Sep 23, 2026
Read in source context →
Continue exploring
prompt caching optimization →

Key passages4

Attributed passages with the context to verify them. Open the original text to check the source.

Paragraph 7

token efficiency optimization

7% token cost reduction without quality loss

Original excerpt

Changes across each of these layers reduced token costs for users by 7% without reducing agent quality.

Jediah Katz, Connor O’Keefe & Calvin Yee·Sep 23, 2026

Context

Over the past few months we've responded to this shift by improving the efficiency of Cursor's agent harness. The harness gives us direct control over how each request is assembled, how context is reused, and when work is divided across agents.

Paragraph 11

system prompt design

66% system prompt reduction enabled by model capability gains

Original excerpt

This was true across model families, allowing us to trim roughly 66% of our system prompt.

Jediah Katz, Connor O’Keefe & Calvin Yee·Sep 23, 2026

Context

As models improved, much of that direction became unnecessary. Instead of long lists of "DO NOT do this," "You must," or "Important" instructions, we could simply define how a tool behaves and models would generally comply.

Paragraph 15

dynamic context loading

46.9% token reduction from dynamic MCP tool loading

Original excerpt

This reduced total tokens by 46.9% across sessions that called an MCP tool.

Jediah Katz, Connor O’Keefe & Calvin Yee·Sep 23, 2026

Context

That created an opportunity to improve efficiency by keeping tools available without including their full definitions in every request. We'd solved a similar problem earlier this year when we moved MCP tools into dynamic context, loading them only when needed.

Paragraph 35

prompt caching optimization

Cache-related changes reduced cold cache misses by 20%

Original excerpt

These changes reduced the rate of cold cache misses by 20%.

Jediah Katz, Connor O’Keefe & Calvin Yee·Sep 23, 2026

Source & methodology

These viewpoints are linked to their original sources. Paraphrases are labeled and are not verbatim quotes.

Open transcript or source material Report an issue