Skip to content
AnthropicAnthropic

Building the future of agents with Claude

Anthropic’s Alex Albert (Claude Relations), Brad Abrams (Product) and Katelyn Lesse (Engineering) discuss the evolution of building agents with Claude, the latest Claude Developer Platform features, and why agents perform best when developers “unhobble” their model with tools. Learn more about the Claude Developer Platform: https://www.claude.com/platform/api 00:00 - Introductions 00:30 - What is the Claude Developer Platform? 2:30 - What is an AI agent 3:15 - Building frontier intelligence for AI agents 4:00 - Reducing model scaffolding to build better agents 5:05 - The evolution of agentic frameworks 6:40 - Unhobbling the model with tools like web fetch 8:35 - Building agents with the Claude Agent SDK (formerly the Claude Code SDK) 10:50 - Best practices for identifying agentic use cases 11: 40 - Driving better agentic outcomes with the SDK 14:35 - Best practices for managing context and memory with Claude 19:00 - The future of the Claude Developer Platform (observability, computer use, and other ways to unhobble the model)

Brad AbramshostAlex AlberthostKatelyn Lessehost
Oct 1, 202522mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

Claude’s agent future: unhobble models with tools and SDKs

  1. The Claude Developer Platform expands beyond an API into a full platform—APIs, SDKs, docs, console, and built-in tools like web search/fetch, code execution, batching, and prompt caching—to help developers build on Claude end-to-end.
  2. Anthropic defines an “agent” as a model that autonomously chooses and uses tools, interprets results, and decides next steps, rather than strictly following a pre-scripted workflow.
  3. As models improve, heavy orchestration and guardrail “scaffolding” can become a liability by constraining intelligence, so the platform focus shifts toward “unhobbling” the model with the right tools instead of rigid frameworks.
  4. The Claude Agent SDK (formerly Claude Code SDK) provides an out-of-the-box agentic loop/runtime that accelerates prototyping and can generalize beyond coding by offering generic capabilities like filesystem access and command execution.
  5. Key near-term platform priorities include better context management (tool-result pruning with tombstones), agent memory primitives, and observability for long-running autonomous tasks, plus a longer-term push toward giving Claude a more persistent “computer.”

IDEAS WORTH REMEMBERING

5 ideas

Treat the platform as more than model access—use built-in tools to unlock capability.

Anthropic positions the Developer Platform as the combination of APIs, SDKs, console workflows, and first-party tools (e.g., web search/fetch, code execution, batching, prompt caching) that collectively raise what agents can do.

An agent is defined by autonomous tool choice and iterative decision-making.

Their practical definition centers on the model selecting tools, calling them, interpreting outputs, and deciding the next action—leaning into model reasoning rather than rigid step-by-step orchestration.

Over-scaffolding can hide improvements from newer models.

Customers sometimes see minimal gains after a model upgrade because their workflow constraints prevent the model from applying its higher intelligence; lighter patterns tend to improve automatically as models advance.

“Unhobbling” means giving the model effective tools, not heavy frameworks.

Instead of opinionated, complex agent frameworks, Anthropic advocates lightweight harnesses plus strong tool primitives—enabling deep research, retrieval, and multi-step work with minimal prompting.

Start with the Claude Agent SDK to avoid rebuilding the agent loop.

The SDK provides a ready-made agentic runtime that automates tool calling and loop management, letting developers prototype quickly and then customize—rather than implementing prompt caching and orchestration from scratch.

WORDS WORTH SAVING

5 quotes

Agents is... It’s almost sort of a buzzword, right?

Brad Abrams

How do you unhobble the model?

Brad Abrams

All we did... we just give it the web search tool, and... deep research tasks are, like, almost completely done.

Brad Abrams

Once we got done removing things... it turns out there was nothing coding left.

Brad Abrams

If we hire an employee... but we don’t give them a computer... they would not be very successful.

Brad Abrams

Claude Developer Platform scope (APIs, SDKs, console, docs)Definition of AI agents and autonomyReducing scaffolding vs. heavy frameworksTooling to “unhobble” models (web search/fetch, code execution)Claude Agent SDK / agentic loop runtimeContext-window management (removing old tool calls, tombstones)Agent memory, observability, and “computer use” roadmap

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome