The original title is "Instrument First, Then Prompt: Finding Real Agentic Pipeline Bugs"

Short summary

When agentic pipelines misbehave, developers instinctively rewrite prompts — but the bug is almost never there. Three common failure modes (chunk boundary truncation, tool output drift, eval dataset blindspots) are invisible in model responses. The author advocates adding instrumentation first: a decorator around tool calls, logging around context injection, and a counter around router branches. Timestamped JSON lines are sufficient to find where data shape mismatches or chunking issues cause silent failures.

• •Prompt rewriting is rarely the fix for agentic pipeline bugs — instrumentation is
• •Three common failure modes: chunk truncation, tool output drift, eval blindspots
• •Practical code for tracing tool calls, context injection, and router branches with plain JSON logs

Generated with AI, which can make mistakes.