Why (Senior) Engineers Struggle to Build AI Agents

Abstraction: Senior engineers' deterministic instincts conflict with probabilistic agent design

Key points:

• Junior engineers often ship agents faster than seniors because seniors try to "code away" the probabilistic nature by forcing strict data structures and control flow
• Five traps: forcing structured types on natural-language state; hard-coding control flow instead of trusting the LLM; crashing on errors instead of feeding errors back as inputs; writing unit tests instead of running evals; and building "human-grade" ambiguous APIs instead of verbose semantic tool definitions
• "Text is the new State": store user preferences and intents in natural language, not booleans — downstream agents can read richer context
• Evals over tests: measure Reliability (pass@k) and Quality (LLM-as-judge) rather than binary assertions; a 45/50 pass rate with 4.5/5 quality can be production-ready
• Tools must be "idiot-proof" with verbose names (user_email_address not email) and descriptive docstrings that serve as prompt engineering for the model
• Agents allow just-in-time API adaptation — unlike code, an agent reads a new tool definition and adjusts without needing a redeployment

Connections: Philipp Schmid · AI Agents · Agent Engineering · Evals

Source: https://www.philschmid.de/why-engineers-struggle-building-agents