Beating AI Flash: Factory releases Coding Agent runtime framework Droid 1.0, introducing programmatic tool calling (Code Mode) and making it the default execution method.AI can write a JavaScript script first, then use the script to make hundreds of consecutive tool calls without requesting the large model again every few steps, thereby reducing the number of model calls and Token consumption.
For example, migrating and verifying 151 Linear issues, the traditional approach is estimated to require 36 model requests and 3100000 input Token. With Code Mode instead, it actually requires only 2 requests and approximately 61000 Token, with input Token reduced by approximately 98%. In broader testing, task costs using Claude Opus 5.5 decreased by approximately 15%.
Factory also optimized tool-calling strategies, caching, and task execution mechanisms for different models. In Terminal-Bench 4 official self-test, using the same GPT-6.1 Sol model, Droid scored 3.5 points higher than Codex, with successful-task costs reduced by 12%, but execution time increased by 10%。
Code Mode is not a new technology. Cloudflare and Anthropic had previously disclosed similar approaches, and Pi, OpenCode 2.0 and DeepSeek Harness also support them.
Original link https://m.theblockbeats.info/flash/374630
Droid 1.0 introducing programmatic tool calling Code Mode, and making it the default execution method.
03
The article states that AI can write a JavaScript script first, then use the script to make hundreds of consecutive tool calls.
04
In an example migrating and verifying 151 Linear issues, the traditional approach is estimated to require 36 model requests and 3100000 input Token；Code Mode the actual use required 2 requests and approximately 61000 Token, with input Token reduced by approximately 98%。
05
The article states that, in broader testing, task costs using Claude Opus 5.5 decreased by approximately 15%。
06
In Terminal-Bench 4 the official self-test, using the same GPT-6.1 Sol model, Droid scored 3.5 points higher than Codex, with successful-task costs reduced by 12%, while execution time increased by 10%。
07
The article states that Cloudflare and Anthropic had previously disclosed similar approaches, and Pi, OpenCode 2.0 and DeepSeek Harness also support them.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
Code Mode allows AI to write a JavaScript script first, then have the script operate multiple tools consecutively, reducing the number of repeated large-model requests during the process. The reported issue example shows that the number of requests and input Token consumption decreased significantly; however, the benchmark results came from an official self-test.
Why it matters to readers
If these test results can be replicated on other tasks, programmatic tool calling may reduce the number of model calls and Token consumption for complex automated tasks, although the report also shows increased execution time.
Beginners can understand it as follows: let AI write an automated workflow in advance, then complete multiple operations at once instead of asking the model again every few steps. The key questions are whether the savings are reproducible and whether the execution time is acceptable.
Risks and unknowns
The article's performance, cost, and benchmark data were not independently replicated.
151 The traditional-approach data in the Linear-issue example are estimates.
It remains unclear whether different tasks, models, and production environments can achieve the same magnitude of Token and cost improvements.
Code Mode The stability, security, and script maintenance costs were not specified in the evidence.
Related Developments
Loading event timeline…
Related concepts
Coding Agent
This term is not in the glossary yet. Browse related concepts in the glossary.
