Architecting Agentic Systems with Engineering Discipline

Systems Thinking and Decomposition

Instead of relying on monolithic prompts that grow into unmanageable blobs, engineers should treat agents as components within a larger system. Decomposition involves identifying distinct tasks—such as data normalization, commute calculation, and neighborhood research—and separating them into modular units. This approach mirrors traditional software design, where logic is divided into services or layers to improve testability and maintainability.

Separation of Concerns and Tooling

Effective agentic systems leverage the right tool for the specific job. Deterministic tasks, such as calculating commute times or deduplicating data, should be handled by standard code rather than LLMs to ensure reliability and lower costs. Agents should be reserved for tasks requiring judgment, ambiguity, or reasoning. By defining clear schemas for agent outputs, developers create contracts that allow downstream processes to consume data reliably, moving beyond the limitations of session-bound text.

State Management and Safety

Agentic workflows must be designed for idempotency to handle failures and retries without causing side effects. Systems should log actions to a persistent memory layer, allowing the agent to check the current state before executing subsequent steps. Furthermore, security engineering principles apply: treat all external data as untrusted input and enforce strict boundaries on agent capabilities, such as requiring human approval for high-stakes actions like booking tours or submitting offers. This reduces the blast radius and ensures the system remains maintainable and predictable.