TL;DR
- Coding agents make decisions. They don’t execute fixed checklists.
- Because they prioritize, they will miss things.
- One review pass is not enough.
- Repeated and focused review prompts dramatically reduce blind spots.
- “Review at least 3 times” is a simple but effective mitigation.
If you skip structured review cycles, you are optimizing for speed over reliability.
The Core Problem
A static script follows instructions. A coding agent interprets them.
That means it:
- Decides what is important.
- Decides what is “probably not required.”
- Compresses reasoning.
- Makes trade-offs.
Those trade-offs are not always correct.
Agents don’t “forget” steps. They deprioritize them.
Why One Review Is Not Enough
An agent review is also probabilistic.
If you ask:
“Review the implementation.”
You get one prioritization pass.
Blind spots remain because:
- Attention is focused on dominant concerns.
- Secondary risks get compressed away.
- Ambiguities are resolved silently.
A second and third pass shifts attention. That alone increases issue coverage significantly.
Repetition reduces blind spots.
Agent Review vs Checklist Review
Checklist / static analysis
- Deterministic.
- Good for compliance and style.
- Cannot reason about architecture or intent.
Agent review
- Reinterprets requirements.
- Evaluates trade-offs.
- Identifies missing abstractions.
- But can miss things for the same reason generation did.
They are complementary. Not interchangeable.
Why Agents Miss Things
Common patterns:
- Over-optimization for the main goal (“feature works”).
- Assumptions about context (“validation handled elsewhere”).
- Silent scope narrowing.
- Defaulting to familiar but slightly outdated patterns.
This is expected behavior in any system that optimizes.
Simple Mitigations That Work
You don’t need complex orchestration.
1. Force multiple passes
Example:
Review at least 3 times.
Each pass must identify new issues.
Do not repeat prior findings.
Focus on different risk categories.
This alone increases depth.
2. Use focused review prompts
Examples:
- “Perform a strict security review.”
- “Perform a modernization review.”
- “Review for architectural consistency.”
- “Review for hidden coupling.”
- “Review for testability gaps.”
You are guiding attention, not increasing intelligence.
3. Separate generation from review
Bad pattern:
- Generate.
- “Looks good?”
Better pattern:
- Generate.
- Critical review.
- Refactor.
- Focused review.
- Finalize.
Hard Truth
If you use agentic coding without repeated review cycles:
- You are trading reliability for speed.
- You are increasing silent technical debt.
- You are relying on optimistic interpretation.
The model is not the risk. The missing review discipline is.
Originally published as an internal Netresearch wiki article in February 2026. Republished here lightly edited for public context.