{
  "schema_version": "1.1",
  "id": "s3:https://cursor.com/blog/improved-token-efficiency",
  "slug": "improved-token-efficiency-1oyiyul",
  "url": "https://feed7.dev/p/improved-token-efficiency-1oyiyul",
  "title": "Improved token efficiency for longer agent runs",
  "why_included": "Cursor cut agent token cost by shrinking static instructions, loading tools on demand, improving cache boundaries, and using subagents more selectively.",
  "summary": "Cursor reports **7% lower user token costs** without reduced agent quality after trimming its system prompt by about **66%**, cutting built-in tool descriptions in static context by **60%**, and reducing cold cache misses by **20%**.",
  "practical_implication": "Harness builders should keep only frequently needed tools in static context, place variable setup after stable cache boundaries, reduce incidental file-read overhead, and delegate when context isolation outweighs coordination cost.",
  "agent_context": "Cursor reports **7% lower user token costs** without reduced agent quality after trimming its system prompt by about **66%**, cutting built-in tool descriptions in static context by **60%**, and reducing cold cache misses by **20%**.\n\nHarness builders should keep only frequently needed tools in static context, place variable setup after stable cache boundaries, reduce incidental file-read overhead, and delegate when context isolation outweighs coordination cost.\n\nThese results come from Cursor production traffic and A/B tests, so they may not transfer directly to another request mix. The post also warns that subagents can duplicate work when isolated contexts lack coordination.",
  "source": {
    "name": "Cursor",
    "url": "https://cursor.com/blog/improved-token-efficiency",
    "published_at": "2026-09-23T12:00:00.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Engineering Post",
  "layer": "agent",
  "domains": [
    "coding"
  ],
  "topics": [
    "harness-engineering",
    "context-caching",
    "subagents"
  ],
  "verification": {
    "status": "official_source",
    "label": "Official Source",
    "method": "source_feed",
    "verified_at": null
  },
  "uncertainty": [
    "These results come from Cursor production traffic and A/B tests, so they may not transfer directly to another request mix. The post also warns that subagents can duplicate work when isolated contexts lack coordination."
  ],
  "connected_context": {
    "meaning": "This adds production A/B evidence that substantial harness savings can come from shrinking static instructions, tool schemas, and cache misses without changing models. It reinforces deliberate compaction and selective context loading, while narrowing the case for subagents: isolation can save context, but poor coordination may duplicate work and erase those gains.",
    "corpus_size": 875,
    "generated_at": "2026-09-25T09:06:52.652Z",
    "connections": [
      {
        "title": "Anthropic's CCA Exam as a Field-Guide for Agentic Engineering — Frank Coyle, UC Berkeley",
        "source_name": "AI Engineer",
        "source_url": "https://www.youtube.com/watch?v=Z-c11pV_uvU",
        "feed7_url": "https://feed7.dev/p/anthropic-s-cca-exam-as-a-field-guide-for-agentic-engineering-frank-coyl-00nf896",
        "reason": "Cursor’s measured reductions provide empirical support for deliberate compaction and narrow subagent scope, while replacing the candidate’s example token trigger with workload-specific optimization evidence."
      },
      {
        "title": "Agent swarms and the new model economics",
        "source_name": "Cursor",
        "source_url": "https://cursor.com/blog/agent-swarm-model-economics",
        "feed7_url": "https://feed7.dev/p/agent-swarm-model-economics-02n8muc",
        "reason": "Both indicate that context topology can matter more than model choice, but the new warning about duplicated work qualifies the swarm result by making coordination quality part of the efficiency calculation."
      },
      {
        "title": "WorldFlowAI/everything-claude-code",
        "source_name": "GitHub",
        "source_url": "https://github.com/WorldFlowAI/everything-claude-code",
        "feed7_url": "https://feed7.dev/p/everything-claude-code-11d9320",
        "reason": "The repository’s warning against loading an entire harness catalog is reinforced by Cursor’s finding that trimming static prompts and built-in tool descriptions reduced user token costs."
      },
      {
        "title": "It’s Tokens All The Way Down: How RLMs are Different — Kevin Madura, AlixPartners",
        "source_name": "AI Engineer",
        "source_url": "https://www.youtube.com/watch?v=xo68uCibfm8",
        "feed7_url": "https://feed7.dev/p/it-s-tokens-all-the-way-down-how-rlms-are-different-kevin-madura-alixpar-1w7pcym",
        "reason": "RLM-style programmatic selection is a complementary response to context cost: Cursor minimizes always-present harness context, while RLMs keep large task inputs outside direct attention and inspect selected portions."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-09-23T12:00:00.000Z",
  "modified_at": "2026-09-23T12:00:00.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/improved-token-efficiency-1oyiyul",
    "json": "https://feed7.dev/p/improved-token-efficiency-1oyiyul.json",
    "markdown": "https://feed7.dev/p/improved-token-efficiency-1oyiyul.md"
  }
}