TechPrufer
Get started
Anthropic
Anthropic logoAnthropic · Feature trail

AI Agent Code Execution

1 post · Sep 2025 – Jun 2026

What this trail covers

Anthropic is developing and refining methods for AI agents to interact with external systems more efficiently. This includes the development of the Model Context Protocol (MCP) as an open standard for connecting agents to tools and data. Recent advancements focus on enabling code execution environments for agents, allowing them to load tools on demand, process data locally before returning it to the model, and manage state across operations. This approach addresses challenges like context window limitations and improves agent performance through iterative tool development and evaluation, including AI-assisted tool optimization and the establishment of key principles for writing effective agent tools such as namespacing, meaningful context return, token efficiency, and clear tool descriptions.

All Anthropic trails
  1. Paving the way for AI agents in biologyThis post highlights the critical need for deterministic retrieval layers and agent-friendly biological data infrastructure to enable reliable AI agent workflows. It demonstrates through a case study using scientific research agents on NCBI Virus that even advanced models struggle with inconsistent data formats and scattered databases. The addition of 'gget virus,' a deterministic retrieval layer, significantly improved accuracy to nearly 100%. The broader lesson is that for agents to be effective in scientific discovery, biological databases must be designed with agents as scaled users, providing structured workflows and reliable interfaces akin to software infrastructure.
  2. Introducing advanced tool use on the Claude Developer PlatformThis post introduces three new features for the Claude Developer Platform: Tool Search Tool for on-demand tool discovery, Programmatic Tool Calling for orchestrating tool usage via code execution, and Tool Use Examples for demonstrating effective tool usage. These features aim to reduce token consumption, improve accuracy, and decrease latency by allowing agents to dynamically load tools, process intermediate results in a code execution environment, and learn usage patterns from examples. The post details the trade-offs and benefits of these features, including significant token savings (e.g., 85% reduction with Tool Search Tool) and improved accuracy benchmarks.
  3. Code execution with MCP: building more efficient AI agentsThis post introduces code execution as a method to improve AI agent efficiency when interacting with MCP servers. It details how presenting MCP servers as code APIs, rather than direct tool calls, reduces token usage by allowing agents to load tools on demand and process data locally. The post highlights benefits such as reduced context window bloat (e.g., 98.7% saving in the example), efficient data filtering and transformation, improved tool composition through familiar code patterns (loops, conditionals), enhanced security via data tokenization, and state persistence by writing intermediate results to files. It also notes the operational overhead and security considerations of running agent-generated code in a sandboxed environment.
  4. Writing effective tools for AI agents—using AI agentsThis post details Anthropic's techniques for improving the effectiveness of tools used by AI agents, focusing on iterative development and evaluation. It introduces principles for writing high-quality tools, including choosing the right tools, namespacing, returning meaningful context, optimizing for token efficiency, and prompt-engineering tool descriptions. The post also describes a workflow for building and testing tool prototypes, running comprehensive evaluations with agents, and collaborating with AI like Claude Code to automatically improve tool performance. It highlights the difference between deterministic software contracts and the contract between deterministic systems and non-deterministic agents, emphasizing the need to design tools specifically for agents.