In this talk, Justin Schroeder from Standard Agents discusses the critical role domain-specific agents will play in the future of AI. He begins by framing the current AI era as akin to an accelerated Industrial Revolution, where instead of harnessing energy with machines, we are now harnessing intelligence with agents. Schroeder defines agents as deterministic software that leverages the non-deterministic outputs of models to achieve specific objectives. Despite the widespread building of custom agents across various industries, there remains a lack of clear understanding and standardization around what agents are and how to build them effectively.
Schroeder highlights the challenges in building robust agents, including the complexity of orchestrating agentic loops, handling provider abstractions, ensuring durable execution, and managing telemetry and observability at scale. He points out that agents are often not portable or composable, making it difficult to reuse or share them across different environments or applications. As a result, many organizations abandon agent development in favor of Model Context Protocols (MCP) or skills, which, while useful, have limitations such as diminishing returns when scaling and insufficient integration capabilities.
The core of Schroeder’s argument is the distinction between inheritance and composition in agent design. Current approaches tend to inflate a single agent’s context with numerous tools and skills (inheritance), which becomes inefficient and unwieldy as the number of capabilities grows. Instead, he advocates for a compositional approach where multiple small, domain-specific agents operate independently with their own context and tools, coordinated by a higher-level agent through natural language communication. This mirrors how complex projects, like the Apollo moon landing, were managed by specialized teams working together, and it offers greater efficiency, scalability, and control.
Domain-specific agents, Schroeder explains, are far more token-efficient, cost-effective, and practical, especially when using smaller language models or specialized non-language models. They allow strict control over capabilities and permissions, addressing security and operational concerns that large, general-purpose agents struggle with. Although these agents are not yet widely available publicly, Schroeder predicts a rapid rise in their adoption and development throughout 2026, with 2027 becoming the year of multi-agent orchestration. This shift is driven by rising token costs and the need for more efficient, scalable AI solutions in business and customer-facing applications.
Finally, Schroeder outlines an ideal agent architecture that includes models, system prompts, tools (functions, prompts, or other agents), hooks for side effects, and agent-specific rules. He emphasizes the importance of sandboxed file systems and code execution environments for safety and functionality. He envisions a layered ecosystem of domain-specific sub-agents—such as Salesforce, Google Workspace, asset generation, and legal compliance agents—that collaborate seamlessly under a coordinating agent. This modular, compositional design promises to unlock the full potential of AI agents by combining specialization, efficiency, and scalability.
