Anthropic Reveals: Core Methodology for Building More Powerful Claude Agents

Anthropic's Alex (Claude Relationship Lead) and Eric (Multi-Agent Researcher) recently had an in-depth conversation, sharing their latest thinking and practical experience in building AI agents. From why Claude excels at agentic tasks to cutting-edge explorations in multi-agent systems, this conversation covered the core topics in current AI Agent development.

Why Claude Excels at Agentic Tasks

During training, Anthropic had Claude extensively practice acting as an agent. Specifically, the team gave Claude open-ended problems, letting it take multiple steps, use tools, explore environments, and ultimately arrive at answers. Through extensive Reinforcement Learning (RL) across different environments like coding tasks and search tasks, Claude gradually mastered the ability to complete complex tasks with minimal guidance.

Reinforcement Learning and Agent Training: Reinforcement learning is a machine learning paradigm that teaches models optimal behavioral strategies through "trial-and-error-feedback" loops. Unlike supervised learning, which relies on labeled data, RL uses reward signals to guide models toward better decisions in complex environments. In the large language model space, RL typically appears in the form of RLHF (Reinforcement Learning from Human Feedback) or Process Reward Models (PRM), helping models learn multi-step reasoning and tool usage. When training Claude's agentic capabilities, Anthropic leveraged RL to have the model iterate repeatedly in real task environments — failing, adjusting, and retrying — thereby internalizing the meta-ability of "how to self-correct when encountering obstacles," rather than merely memorizing fixed problem-solving patterns.

Eric particularly emphasized the foundational importance of coding ability: "Once you have a great coding agent, it can do any other type of work." Need to search? You can perform web searches via APIs. Need to plan a weekend? You can create a schedule. Coding is seen as the most fundamental skill for agents, with powerful spillover effects — train the hardest thing first, and everything else becomes easy.

An interesting example: Claude can generate Excel spreadsheets by writing Python scripts, or produce complex SVG charts through code. Compared to creating these files directly, the code-based approach is far more efficient than manual click-and-drag operations when handling repetitive patterns. As Eric put it: "Claude gained the power of the for loop."

Claude Code SDK: The Core Tool for Developers Building Agents

Anthropic's Claude Code SDK is becoming the go-to solution for developers building AI agents. Previously, if you wanted to build a coding agent or any type of Agent, you had to start from scratch — calling API endpoints, building loops, building tools, executing tools, interacting with files, and interacting with MCP. Now, all of these capabilities are built into Claude Code.

Technical Background on the MCP Protocol: MCP (Model Context Protocol) is a standardized protocol open-sourced by Anthropic in late 2024, designed to solve the fragmentation problem of integrating AI models with external tools and data sources. Before MCP, every developer had to write custom API adaptation layers for their tools, resulting in massive duplication of effort. MCP defines a unified "tool description-invocation-response" interface specification, similar to what the USB interface does for hardware peripherals — any tool that follows the MCP specification can be directly invoked by MCP-supporting models without additional adaptation. This allows the tool ecosystem to expand rapidly like an app store, letting developers focus on business logic itself rather than the implementation details of underlying communication protocols.

Eric encourages developers to use this SDK as the core of their agent loop: "This way they don't have to spend a lot of time reinventing wheels that we've already spent a lot of time polishing and refining. Instead, they can use it and then add their own custom business logic tools via MCP."

Despite being called "Claude Code," it's actually a general-purpose agent framework. Eric even uses it to plan dates — doing web searches, finding interesting activities and restaurants, completely unrelated to code.

Skills: A More Powerful Capability Injection Mechanism Than Claude.md

Claude.md files let developers define information like coding style and directory layouts. The newly introduced Skills concept goes a step further — you can provide Claude with any type of file: PowerPoint templates, helper scripts, image assets, and more.

Eric used a brilliant analogy: "It's like in The Matrix when Neo first learns kung fu — they inject the kung fu information into his brain and he suddenly becomes a kung fu master." Giving Claude a skill file works the same way — for example, tell it how to create spreadsheets, and it suddenly acts like a banker building financial models for you.

The Evolution from Workflows to Agent Loops

Just a few months ago, the industry was still in the transition phase from workflows to single-agent systems. Now, agent loops have significantly surpassed traditional workflows in absolute quality. Claude has become exceptionally good at responding to feedback and correcting its own work.

The most noteworthy direction of evolution right now is "workflows of agents." Previously, an application might have a workflow where Claude generates a SQL command in a single pass to load data, then passes it to the next step to generate a chart. If the SQL command failed, all subsequent steps would completely collapse.

Now, each step in the workflow is a closed loop — Claude doesn't just write a SQL query once; it runs the query, examines the output, and continues iterating until it confirms the correct values, only then transitioning to the next step. This is a fundamental shift from "chaining prompts" to "chaining agent loops."

Multi-Agent Systems: Making Multiple Claudes Work Together

Multi-agent is fundamentally different from "workflows of agents." The latter is one agent completing its work and passing output to the next; multi-agent involves multiple Claude instances working simultaneously — a parent agent might delegate tasks to five child agents for parallel processing.

Architecture Patterns of Multi-Agent Systems: Multi-agent systems in AI have two mainstream architecture patterns: first, the "Orchestrator-Worker" pattern, where a primary agent handles task decomposition and result aggregation while multiple child agents execute specific subtasks in parallel; second, the "Peer-to-Peer" pattern, where multiple agents communicate and negotiate as equals. Anthropic currently primarily adopts the former. The core challenge of this architecture lies in balancing "context isolation" with "information synchronization" — child agents need sufficient context to complete tasks, but excessive context passing incurs the dual costs of token consumption and latency. This is also why Claude needs to be specifically trained in the meta-skill of "how to give clear instructions to child agents."

Anthropic's deep research search product works exactly this way: the main orchestration agent creates multiple child agents to execute searches in parallel, allowing users to get results faster. In Claude Code, the model also uses child agents — if a subtask requires tens of thousands of tokens (such as finding the implementation of a certain class) but the answer is brief, it can complete this work in a child agent, protecting the main context from being polluted by unnecessary tokens.

Claude's Growth Journey as a "First-Time Manager"

Eric revealed an interesting finding: Claude makes many "first-time manager" mistakes when managing child agents.