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

Source: Dev.to

Overview

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

In this video, Cal from Anthropic’s Applied AI team shares key learnings from building AI agents in 2024, focusing on Claude’s evolution from basic chatbots to sophisticated coding agents. He introduces context engineering as the successor to prompt engineering, explains how agents differ from workflows through their ability to run in loops with tools, and discusses Claude Opus 4.5’s achievements (e.g., 80 % on SWE‑bench). Practical insights cover system prompts, tool design, progressive disclosure through “skills,” handling long‑horizon tasks via compaction and sub‑agents, and the Claude Agent SDK for building custom agents across domains beyond software engineering.

This article is auto‑generated while preserving the original presentation content; minor typos or inaccuracies may be present.

Main Part

Introduction: Cal’s Journey at Anthropic and the Rise of Claude 3

“Alright, hello everybody, let’s get started… If you remember that you had originally signed up for this talk and it was something about Anthropic and Lovable, you are in the right place.”

Cal explains his role and the early days of Anthropic’s Applied AI team:

• Joined Anthropic two years ago to help customers and partners build products on top of Claude.
• Initially worked with Claude 2.1, which had a 200 k token context window (far larger than the 32–64 k windows of competing models) and was available on AWS Bedrock.
• Six weeks later, Anthropic released the Claude 3 family (Opus, Sonnet, Haiku). Claude 3 Opus quickly became a frontier model, dramatically increasing customer interest.

From Q&A Chatbots to Claude Code: The Evolution of AI Applications

Early 2024 – Q&A Chatbots & RAG

• Built Q&A chatbots that retrieved help‑center articles, inserted them into prompts, and asked Claude to answer user questions.

Claude Sonnet 3.5 – A Turning Point

• Claude Sonnet 3.5 positioned as a middle‑tier model: balanced speed, cost, and performance.
• Noticed strong ability to generate HTML, JavaScript, and CSS.
• Created the Artifacts product: when Claude writes an HTML file, the system captures it, opens a side panel, and renders the result.
• Limitations: artifacts required full file regeneration for edits (e.g., changing game scoring required rewriting the entire HTML).

Claude CLI & Claude Code

• Discovered the Claude CLI tool, which allowed rapid prototyping.
• Using Claude Sonnet via CLI, Cal generated a full Angular note‑taking app overnight without writing a single line of code.
• Inspired collaboration with Kat and Boris, founding members of the Claude Code team, to further explore AI‑driven development workflows.

The transcript cuts off here, but the discussion continues with deeper dives into system prompts, tool design, progressive disclosure (“skills”), long‑horizon task handling, and the Claude Agent SDK.