Author(s): Aashrith Sai

Originally published on Towards AI.

Your Agent Isn’t Broken. Your Architecture Is.

Most agent tutorials show you a happy path.

The article argues that agent failures in production usually come from architecture and orchestration choices rather than the model itself. After showing a simple single-agent “happy path” (looping between two tools: parsing resumes and scoring candidates), it demonstrates how adding more requirements (extra sources like LinkedIn/GitHub, employment verification, and human approval for rejections) quickly reveals brittle design: single-agent tool ordering becomes non-deterministic with many tools, prompting the need for a multi-agent coordinator with scoped tools, explicit context injection, and deterministic delegation/aggregation. It further distinguishes guidance vs guarantees—using prompts for soft preferences and “hooks” (hard intercepts) for compliance or legal/financial constraints—and shows that expanding tool access without scoping can let agents bypass required steps (e.g., rejecting directly). Finally, it covers production robustness by treating errors as explicit structured messages (so silence isn’t mistaken for success) and using manifest-based crash recovery to prevent losing state or issuing offers after verification failures, concluding with a decision checklist tying requirements to these architectural layers.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI