AWS re:Invent 2025 - What Anthropic Learned Building AI Agents in 2025 (AIM277)

Summary of "What Anthropic Learned Building AI Agents in 2025"

Introduction

• Presenter is Cal, who joined Anthropic 2 years ago to help start the Applied AI team
• Anthropic's mission is to help customers and partners build great products and features on top of cloud
• Anthropic has evolved from their early model Claude 2.1 to the more advanced Claude 3 model family

Anthropic's AI Safety Research

• Anthropic was founded with a belief that transformational AI would happen faster than expected
• Their AI safety research focuses on two key areas:
  1. Alignment - Ensuring the AI model reflects important values and can identify misaligned behavior
  2. Interpretability - Understanding how the model works under the hood to enable safety and control

Anthropic's Enterprise Focus

• Anthropic has chosen to focus on the enterprise market
• Key benefits include reducing hallucination and the model's willingness to say "I don't know"
• Anthropic has developed solutions like the Model Context Protocol to integrate data from silos at scale

Model Improvements Over Time

• Anthropic has seen significant improvements in their models, with the latest Opus 4.5 model outperforming previous versions
• Opus 4.5 can achieve better results than Sonnet 4.5 with fewer tokens, which is important for cost and latency
• Anthropic has also made progress on mitigating prompt injection attacks, though this remains an ongoing challenge

Future Model Capabilities

• Anthropic expects to see continued improvements in areas like:
  1. Long-running agents that can maintain coherence and progress for days or weeks
  2. Agents that can interact with the web and computer interfaces like a human user
  3. Expanding into new verticals like cybersecurity and financial services analysis

The Importance of Context Engineering

• Moving beyond single-turn prompts and workflows, Anthropic has focused on building "agents" - models running in a loop with access to tools
• Key aspects of effective context engineering include:
  1. Providing minimal yet sufficient instructions in the system prompt
  2. Carefully designing the tools the agent can use and how they are described
  3. Handling long-running tasks by compacting context, allowing the agent to take notes, or using sub-agent architectures

The Cloud Agent SDK

• Anthropic has developed the Cloud Agent SDK to provide a battle-tested foundation for building powerful agent-based applications
• The SDK provides primitives like memory, web search, and orchestration to enable agents to interact with computers and solve a wide range of problems
• Anthropic sees this as a key trend for 2026, with agents gaining access to computer interfaces to tackle tasks beyond just software engineering

Key Takeaways

• Anthropic has made significant progress in building advanced AI models and agents, with the latest Opus 4.5 model outperforming previous versions
• Context engineering is critical for building effective agents, requiring careful design of prompts, tools, and long-running task management
• Anthropic's Cloud Agent SDK provides a robust foundation for building agent-based applications that can interact with computers to solve a wide range of enterprise problems
• Anthropic expects agents with computer access to be a major trend in 2026, enabling them to tackle tasks in verticals like cybersecurity, finance, and beyond