Building Blocks for Tomorrow’s AI Agents

Key Concepts

• Agentic Components: Reusable tools and services for building AI agents.
• Code Execution: A tool that allows Claude to write and execute code to solve problems.
• Web Search: A tool that allows Claude to search the web for information to augment its knowledge.
• MCP Connector: A tool that allows Claude to connect to and use Modular Component Platforms (MCPs) to perform tasks.
• Prompt Caching: A technique to reuse prompts and responses to reduce latency and cost.
• Batch Processing: A method to process large amounts of data asynchronously.
• Priority Tier: Dedicated capacity for reliable and performant service.

Build: Code Execution Tool

• Problem: LLMs are good but sometimes need traditional software development for tasks like advanced data analytics.
• Solution: Code Execution tool allows Claude to write and execute code.
• How it works:
  1. Client sends a prompt to Claude, specifying a container ID.
  2. Claude decides to use the code execution tool.
  3. Claude writes Python code to answer the question.
  4. The code is executed in a dedicated container.
  5. Standard output, standard error, and any created files are returned to Claude.
  6. Claude reasons over the results and provides an answer.
• Example:
  • Prompt: "What is the 100th prime number?"
  • Claude writes code to calculate the 100th prime number.
  • The code is executed, and the result (541) is returned.
  • Claude provides the answer with a joke.
• Implementation: Simple to set up using the messages API with a new "tools" block.
• Shopify Case Study: Using code execution for sidekick agents to help merchants build storefronts and AB testing experiences.
• Demo:
  • Simple query: Calculates the 100th prime number.
  • Complex query: Analyzes uploaded AB test results, calculates statistical significance, and makes a recommendation. Claude analyzes the spreadsheet, writes code, executes it, and iterates to get the desired insights.
• Key Features:
  • Enthropic-hosted computing environment.
  • Flexible and developer-controlled container allocation.
  • Isolated containers for each organization.
  • 50 free container hours to start.
  • Scalable pricing.

Connect: Bringing Data to the Model

Web Search

• Problem: Models are trained on data up to a certain point in time and may lack real-time information.
• Solution: Web search tool allows Claude to access and use current information.
• Use Cases: Financial data (stock prices), legal case law, API documentation.
• How it works:
  1. Client sends a prompt to Claude.
  2. Claude reasons about the task and formulates search queries.
  3. Claude issues queries to a search engine.
  4. Search results (title, URL, content) are returned to Claude.
  5. Claude reasons over the results and may issue additional queries.
  6. Claude produces a report with citations for each fact.
• Implementation: Messages API with a tool similar to code execution.
• Features:
  • Domain restriction for accurate answers.
  • Control over the maximum number of turns.
• Kora Case Study: Using web search for a consumer agent that requires up-to-date information.
• Demo:
  • Query: "What are the bench scores for all of Anthropic's models since 3.5?"
  • Claude searches the web for the information and provides the results.
  • Combined query: "How many elephants can travel over the Golden Gate Bridge in an hour?" Claude uses web search to find the weight capacity of the bridge and the walking speed of an elephant, then uses code execution to calculate the answer.
• Key Features:
  • Anthropic's agentic search capability.
  • Model decides how to search and how many times to search.
  • Citation support for grounded and auditable results.
  • Composable and developer-controlled.
  • Reasonably priced.

MCP Connector

• Problem: Customers want to leverage the growing ecosystem of Modular Component Platforms (MCPs) within their agents.
• Solution: MCP Connector allows Claude to connect to and use MCPs to perform tasks.
• How it works:
  1. Client sends a prompt to Claude.
  2. Claude identifies the necessary MCPs and tools.
  3. Claude makes API calls to the MCPs, potentially using OAuth for authentication.
  4. MCPs perform the requested tasks and return the results to Claude.
  5. Claude combines the results and provides a final answer.
• Example:
  • Prompt: "Create an email with a creative and motivational image about my Asana project status and send it to the team."
  • Claude uses the Asana MCP to get project status.
  • Claude uses an image creation MCP (hosted on Cloudflare) to generate an image.
  • Claude uses the Zapier MCP to send the email.
• Implementation: Messages API with a new "MCP servers" attribute.
• Features:
  • Simple to set up.
  • OAuth support.
  • Standard token prices.
• Zapier Case Study: Mutual customers can build powerful agents by combining Anthropic's MCP API support with Zapier's MCPs.
• Demo:
  • Query: "What are my open tasks in Asana?"
  • Combined query: "Create an email with a creative and motivational image about my Asana project status including some analysis on the percentage complete and any news on the web about those tasks and send it to the team." Claude uses Asana MCP, code execution, web search, and an image creation MCP to complete the task.
• Remote MCPs Available: Task management, payments, video creation, machine management.

Optimize: Improving Performance and Cost

• Prompt Caching:
  • Extends the cache duration to one hour with a 90% discount on cache hits.
• Batch Processing:
  • Supports web search, code execution, and MCP connector.
  • Offers a 50% discount.
  • Enables building async agents quickly.
• Priority Tier:
  • Offers dedicated capacity for reliable service.
  • Provides a discount for longer commitments.
  • Guarantees 99% reliability.

Conclusion

Anthropic's agentic components provide a powerful and flexible platform for building AI agents. The code execution tool, web search, and MCP connector enable agents to access and use a wide range of information and services. Optimization features like prompt caching, batch processing, and priority tier help to improve performance and reduce costs. These tools empower developers to create sophisticated and reliable AI agents for a variety of use cases.