The Failure of Monolithic Inheritance

Modern agent development is currently stuck in an 'inheritance' trap. Developers build large, general-purpose agents (like Claude or ChatGPT) and attempt to extend their capabilities by stacking endless tools, skills, and context (via protocols like MCP). This approach leads to diminishing returns: as the context window inflates with unrelated data, the agent's performance degrades, costs skyrocket, and debugging becomes a nightmare. These monolithic agents are difficult to port, impossible to share, and prone to 'hallucinating' permissions because they are given too much power over too many domains.

The Case for Composition over Inheritance

Schroeder advocates for applying the classic software engineering principle of 'composition over inheritance' to AI. Instead of one massive agent that knows everything, developers should build small, domain-specific agents—each with its own targeted system prompt, minimal toolset, and isolated message history.

These agents act as specialized experts. A 'Figma agent' only knows Figma; a 'Gmail agent' only knows Gmail. A primary coordinator agent orchestrates these specialists using natural language. This mimics the human organizational structure seen in complex engineering feats like the Apollo 11 mission, where teams of experts with specific, limited responsibilities collaborated to achieve a larger goal.

Why Domain-Specific Agents Win

  • Token Efficiency: By isolating context, agents only process the information relevant to their specific task, often achieving over 80% greater token efficiency.
  • Cost Optimization: Small, specialized tasks can be offloaded to smaller, cheaper models (e.g., DeepSeek V4 Flash) rather than relying on expensive frontier models for every step. This is critical for customer-facing products where unit economics matter.
  • Reliability and Security: Smaller agents have strictly defined capabilities. You can enforce granular permissions, ensuring an agent only accesses what it absolutely needs, which significantly reduces the risk of 'flying too close to the sun' with over-privileged models.
  • Scalability: Because each agent is an isolated execution environment, they can be parallelized and deployed across distributed infrastructure without requiring massive, monolithic compute clusters.

The Path Forward

Schroeder predicts that 2027 will be the year of 'multi-agent orchestration.' As token costs rise (he notes a 30% increase in cost-per-IQ-adjusted-token in 2026), businesses will be forced to move away from inefficient, monolithic agent designs. The emergence of frameworks like Vercel's 'Eve' signals the beginning of this shift toward modular, composable agent ecosystems.

Key Takeaways

  • Stop building monolithic agents: Avoid the temptation to shove every tool and skill into a single system prompt.
  • Adopt composition: Break complex workflows into smaller, domain-specific agents that communicate with a central coordinator.
  • Prioritize portability: Design agents that are self-contained, making them easier to test, share, and deploy across different environments.
  • Optimize for cost: Use smaller models for specialized tasks to keep inference costs sustainable for production-scale applications.
  • Focus on observability: Small, isolated agents are easier to trace and debug than large, non-deterministic monolithic systems.