The provided paper introduces LangSmith Agent Builder, a no-code platform designed to enable anyone to create production-ready AI agents, distinguishing them from traditional AI workflows. The core methodology of Agent Builder centers on a conversational, agent-assisted approach to building other agents, departing from deterministic, "if-this-then-that" workflow builders.

Traditional workflows are described as requiring users to pre-define a fixed, step-by-step path for tasks, with small components delegated to Large Language Models (LLMs), necessitating extensive upfront consideration of edge cases. In contrast, LangSmith Agents are dynamic; they reason on the fly, adapt to new information, and autonomously determine the appropriate steps, including creating plans and delegating work to subagents. A key technical distinction is their ability to work on tasks until completion by calling tools in a loop. This allows for iterative processes, such as multiple searches across different tools, synthesis of findings, and continuous operation. Agents also improve over time through user feedback, leveraging short-term memory (conversation context) and long-term memory (to capture preferences and feedback, stored as guidance in the system prompt).

The LangSmith Agent Builder itself operates as an agent, guiding the user through a chat interface from initial idea to a deployed agent. This "agent building agent" applies best practices learned from large-scale agent deployments, abstracting complex agent engineering principles. For example, it automatically generates detailed and lengthy prompts for the user's agent and assists in selecting necessary tools. The platform facilitates the integration of "bring your own tools" via an MCP (Managed Computing Platform) server, allowing connection to external APIs and internal systems, with governance over approved capabilities.

Key features and methodological aspects of the beta release include:

• Chat-based Agent Creation: Users interact with the Agent Builder via natural language, describing the desired agent behavior.
• Automated Prompt Engineering: The Agent Builder generates robust system prompts, often several paragraphs long, based on user input, encapsulating the agent's instructions and behavior.
• Dynamic Tool Chaining and Looping: Unlike fixed workflows, agents can call tools, evaluate results, and iteratively re-engage tools until a task is deemed complete. This enables complex, multi-step research and analysis.
• Memory Integration: Agents utilize conversational memory for short-term context and persist user feedback and preferences into long-term memory (specifically, the system prompt) to continuously refine their performance.
• Tool Management and Governance: Technical teams can securely expose internal tools via MCP servers. Non-technical users can then select from these approved tools for their agents, authenticating via OAuth without direct IT intervention.
• Workspace Collaboration: The platform introduces workspaces allowing teams to browse, copy, and customize agents, promoting reusability and collaboration via one-click cloning.
• Multi-model Support: Users can choose between different LLM providers (e.g., OpenAI, Anthropic) based on their specific task requirements.
• Programmatic Invocation: Agents can be invoked via API, allowing them to be embedded into existing software systems and workflows.
• Natural Language Updates: Agents can be modified or extended by simply telling the Agent Builder what to change, which updates the agent's system prompt without requiring a rebuild.

The platform aims to address two common challenges in AI adoption: the difficulty for non-technical users to build effective AI tools (by providing an intuitive, prompt-less interface for agent creation) and the need for technical teams to balance speed with security and governance (by offering centralized tool management and secure access). Examples of agents built by users include role-specific research agents (e.g., sales prospect research, market research), agents for turning ambient information into tracked projects (e.g., creating Linear issues from Slack messages), and communication/time-saving assistants (e.g., email triage, calendar management). The underlying methodology empowers users to operate like "managers" describing desired outcomes, with the agent autonomously figuring out the execution details and iteratively working towards task completion.