Building the future of agents with Claude

Detailed Summary

Introductions and Platform Overview (0:00 - 2:29)

• Alex Albert (Claude Relations), Brad Abrams (Product), and Katelyn Lesse (Engineering) introduce themselves and the topic of building agents with Claude.
• The Anthropic API has been rebranded as the Claude Developer Platform to reflect its expanded capabilities, including APIs, SDKs, documentation, and console experiences.
• The platform supports customers in raising the ceiling of intelligence with Claude and is also used internally for products like Claude Code.
• New features added over the past year include prompt caching, a batch API, web search/fetch, context management, and code execution, transforming it into a comprehensive platform.

Defining AI Agents and Reducing Scaffolding (2:30 - 5:04)

• An AI agent is defined as a model that autonomously chooses and calls tools, handles results, and determines next steps, emphasizing the model's reasoning.
• While predefined workflows are useful, agentic patterns allow services to improve automatically with new model releases.
• Heavy scaffolding can limit a model's intelligence, preventing it from fully utilizing new capabilities in updated models.
• As model intelligence increases, less scaffolding is needed, allowing the model to have a greater contextual understanding of tasks and reducing the need for guard rails.

Evolution of Agentic Frameworks and Unhobbling the Model (5:05 - 8:34)

• There's ongoing discussion about what constitutes an agent, with some arguing for simple while loops, while others advocate for more complex frameworks.
• Many existing frameworks have become too heavy and opinionated, hindering the model's performance.
• The strategy is to provide auxiliary tools and frameworks that are opinionated on how to use tools but remain lightweight, avoiding overly heavy structures.
• The concept of "unhobbling" the model means giving it the necessary tools and freedom to leverage its inherent capabilities.
• An example is the server-side web search and web fetch tools, which enable Claude to perform deep research tasks autonomously by calling tools, analyzing results, and performing further searches.
• This shifts intelligence from the developer's explicit guidance to the model's autonomous problem-solving.

Building Agents with the Claude Agent SDK (8:35 - 11:39)

• The Claude Agent SDK (formerly Claude Code SDK) is recommended for developers starting to build agents.
• It functions as a general-purpose agentic harness, automating the loop of tool calling and feature use.
• Initially built for coding, it was quickly realized to be an excellent general-purpose agentic harness, abstracting away the need to build complex loops and tool management.
• The SDK is built on the messages API and existing tools, providing a higher level of abstraction for developers.
• Identifying agentic use cases involves focusing on business value, such as saving engineering hours or reducing manual work, to define the project's scope.

Driving Better Agentic Outcomes with the SDK (11:40 - 14:34)

• The SDK is suitable for enterprises, allowing deployment of the agentic loop runtime wherever needed.
• The goal is to provide higher-order abstractions that not only simplify development but also ensure the best outcomes by leveraging Anthropic's internal knowledge of research and inference.
• A key focus for the future is observability for long-running tasks, enabling developers to monitor, steer, and tune agents effectively.
• Observability is crucial for auditing autonomous systems and ensuring they perform as intended, allowing for adjustments and improvements.

Best Practices for Managing Context and Memory (14:35 - 18:59)

• Web search and web fetch are important tools for agents.
• Managing the context window is critical; while Claude has a 200k token default and a 1 million token beta on Sonnet, smaller, focused contexts often yield better results.
• A new feature allows the model to remove older, unneeded tool calls from the context, similar to decluttering a desk, to improve focus.
• Guard rails are in place to prevent the removal of necessary context, preserving recent tool calls and using "tombstones" to indicate removed information without completely wiping the model's memory.
• An agentic memory feature allows the model to take notes during tasks, learning and improving over time, much like a human would.
• Developers currently manage where this memory is stored, providing control over data storage.

The Future of the Claude Developer Platform (19:00 - 22:05)

• Future plans include higher-order abstractions combined with observability to create a "flywheel" effect, leading to self-improving and continuously enhancing outcomes from Claude.
• This aims to provide "aha moments" where Claude's performance consistently improves over time.
• Brad Abrams expresses excitement for new model launches and the new use cases they unlock.
• A significant future development is giving Claude a "computer" – a persistent environment where it can organize files, set up tools, and execute code (building on the current code execution tool).
• This will further "unhobble" the model, allowing it to perform more complex tasks like advanced data analysis with charts and graphs.

Notable Quotes

"Because as a developer like my creativity ends at some point. I can only think of so many use cases. But the model like anything anything comes with the model will figure out a way to go do that thing."
— Brad Abrams

"In anthropic, what we really think about agent is where the model is taking some uh autonomy to be able to choose what tools to call to call those tools to handle the results and kind of choose the next step."
— Brad Abrams

"We think about we think about it as like how do you unhobble the model like the model already has a lot of capabilities in fact I'm convinced that even if you take a current generation of models there's way more intelligence in there than we've been able to unlock"
— Brad Abrams

"It's like once we got done removing things from the from cloud code to really unhobble the model, it turns out there was nothing coding left."
— Brad Abrams

"If the entire goal here is to help our users like really raise that ceiling of intelligence, get the absolute best outcome out of the models, then higher order abstractions are not just make it easier because you don't have to write all that code yourself. It's actually like how can we like really truly help you get the best outcome"
— Katelyn Lesse

"I'm really excited about giving Claude a computer and I you see like the very baby steps of that with the code execution tool"
— Brad Abrams