Anthropic: Rebuilding from SaaS to agent

• Retool transitioned from a low-code, drag-and-drop platform to an AI-native product called R2. — ✓ “From SaaS → Agent-Native”
• The previous AI assistant, Assist, failed due to quality ceilings imposed by proprietary components and reliability issues. — ✓ “Assist could simultaneously produce these incredible applications”
• Teaching models complex properties like the 155 attributes of a table component was too difficult for finite context windows. — ✓ “table component, which had 155 configurable properties”
• The new architecture uses a 'prompt first' form factor where users type requests into a prompt box. — ✓ “you want to give people an entry point into your world via a prompt”
• R2 generates React and TypeScript code directly instead of using Retool's proprietary Domain Specific Language (DSL). — ✓ “Effectively, it's prompt in, React and TypeScript out”
• Retool preserved its data connection layer, governance model, and SDLC workflows while discarding the component model. — ✓ “Our component model: throw it away”
• Abstractions were made code-first with TypeScript client libraries so agents can interact with data sources. — ✓ “built an abstraction layer on top that's TypeScript-first”
• A small team built the prototype starting on Christmas Eve using Opus to ensure speed and alignment. — ✓ “coding started on Christmas Eve using Opus”
• Internal adoption efforts included providing token budgets and lunch-and-learns to build team intuition. — ✓ “We gave people token budgets”
• R2 resulted in an 11 to 12 times increase in customer engagement and a 75 to 80 percent reduction in production errors. — ✓ “increased by around an order of magnitude – roughly 11 to 12 times”
• The rollout strategy involved design partners, aggressive internal marketing, and separate handling for enterprise deployment. — ✓ “marketed it aggressively, both inside and outside of the product”
• Users shifted from simple Q&A tasks to building full-stack applications and running autonomous work for hours. — ✓ “asking it to build entire full-stack applications”

• 读了 31 张图，每张都有描述
• 图上文字 290 行
• 没读到：2 of 31 images skipped (the download or the vision read failed)
• 没读到：comments not read

样例，第 1 张图 — A presentation slide on a large screen at an event. The slide has a pink background and displays the title 'From SaaS → Agent-Native' along with the subtitle 'How we re-built the product, team, and technology to transition into an AI world.' under the ANTHROPIC logo. 图上文字开头： “ANTHROPIC From SaaS → Agent-Native How we re-built the product, team, and techno…”