### Investigation Summary: "Code Mode" Token Savings and Efficiency #### What Changed? "Code Mode" (popularized by Cloudflare in late 2025/early 2026 and reinforced by Anthropic's Model Context Protocol (MCP) engineering guides) marks a fundamental shift away from **direct tool calling** toward **code execution and sandboxed script writing**. - **The Traditional Approach (Direct Tool Calls):** AI agents load hundreds or thousands of individual tool definitions upfront directly into their context window. When performing batch or multi-step tasks (e.g., scanning 24 workflows and 60 schedules with 26 underlying API actions), every intermediate JSON payload and tool result passes directly through the model's context. This bloats context windows, balloons costs, and increases latency. - **The "Code Mode" Approach:** Instead of exposing thousands of raw endpoints or requiring N individual sequential turns, the agent is given a sandboxed execution environment (like a file tree of typed APIs/scripts or a script runner). The agent writes a single script to orchestrate calls internally, filtering and distilling data *inside* the sandbox. Only the final distilled result (e.g., a single markdown summary or JSON object) reaches the model’s context. #### Why It Matters 1. **Massive Token & Cost Reductions:** - **Agent Swarm (Production Metrics):** A workflow health triage task requiring 26 separate calls dropped from an estimated **~815,000 raw tokens (~$2.44+ floor)** down to **~6,450 tokens (~$0.02)** — achieving a **99.2% cost and character reduction**. - **Cloudflare:** Compressed an expansive API with over 2,500 endpoints from **1.17 million tokens down to roughly 1,000 tokens** (a ~99.9% reduction), solving the problem of tool definitions exceeding LLM context windows. - **Anthropic:** Documented similar reductions (e.g., tasks dropping from 150,000 tokens to 2,000 tokens, or ~98.7% savings) by avoiding duplicate round-trips of large data payloads (like long transcripts or extensive datasets) through the model context. 2. **Eliminates Context Overload & Error:** By preventing intermediate raw payloads from accumulating across lengthy conversational turns, agents maintain a clean context window, avoid hitting token limits, and make fewer data-copying mistakes. #### Strongest Sources Citing the Trend 1. **Agent Swarm Blog:** [*26 Tool Calls, One Script, $0.02: Measuring “Code Mode” in Production*](https://www.agent-swarm.dev/blog/code-mode-token-savings) — Live production benchmarks measuring real workflow automation scans. 2. **Cloudflare Blog:** [*Code Mode: give agents an entire API in 1,000 tokens* / *Code Mode: the better way to use MCP*](https://blog.cloudflare.com/code-mode-mcp/) — Details how 2,500+ endpoints were condensed into a fixed footprint. 3. **Anthropic Engineering:** [*Code execution with MCP: Building more efficient agents*](https://www.anthropic.com/engineering/code-execution-with-mcp) — Explores how progressive disclosure and code execution environments solve context window bloat in the Model Context Protocol standard.