Descriptions:
At the AI Dev 25 NYC conference hosted by DeepLearning.AI, Anthropic engineers Tanveer Mittal (Applied AI team) and Utkarsh Lamba (Infrastructure) delivered a detailed overview of Anthropic’s developer platform and the Claude Agent SDK. The talk covers the full product stack—from the Haiku, Sonnet, and Opus model tiers at the base layer, through agentic capabilities like memory, web search, and multi-agent orchestration, up to consumer products like Claude.ai and Claude Code.
A significant portion focuses on memory systems for long-running agents: the challenge of infinite context, how file-system-based memory allows agents to write and query structured notes across sessions, and how Claude Sonnet 4.5 has begun organizing its own memory more efficiently without explicit prompting. The Claude Plays Pokémon Twitch stream (twitch.tv/claudeplayspokemon) is highlighted as a live eval that has surfaced meaningful research insights—earlier models created a single monolithic markdown file of notes while Sonnet 4.5 models spontaneously adopted per-town, per-gym structured files.
The presentation also covers Claude Sonnet 4.5’s improved computer use capabilities (screenshots, browser navigation, real-world actions), prompt caching as a token and latency optimization, and the Model Context Protocol (MCP) registry—which Anthropic invented—as a core strategy for grounding Claude in application-specific knowledge. Deep partnerships with AWS Bedrock and Google Cloud Vertex AI are noted for teams deploying on those platforms.
📺 Source: DeepLearningAI · Published December 05, 2025
🏷️ Format: Keynote Launch
