AI agent skips requirements/test cases and writes code without approval — causes cascading failures

[tzb1-ai]

GitHub Issue: AI Agent Repeatedly Violates Requirements-First Process

Title

AI agent skips requirements/test cases and writes code without approval — causes cascading failures

Labels

process, incident, ai-tooling

Description

The AI coding agent (Kiro) repeatedly violates the established workflow:

1. Requirements document
2. Test cases
3. User approval
4. Implementation

Instead it jumps directly to code, introduces bugs, then spends hours debugging issues that would have been caught by step 1.

Examples from Aug 14–20

Date	Violation	Impact
Aug 14	NAT instance terraform — no requirements, used yum on AL2023	3+ hours debugging, instance never worked
Aug 14	Scanner scan.sh — wrong git auth format, used /tmp	All clones failed, multiple rebuild cycles
Aug 15	VPC endpoints — deployed SSM only, missed ECR/S3/DynamoDB	Scanner couldn't pull image
Aug 20	Missing security group on ECS task	Task stuck in PENDING, another hour wasted
Aug 20	Asked user "do you want requirements?" instead of defaulting to them	User had to correct again
Multiple	Gave raw copy-paste commands instead of scripts	Not repeatable, not logged, not auditable

Root Cause

Agent prioritizes speed over correctness. It tries to "complete fast" by skipping process steps, which paradoxically makes everything take 5–10x longer due to cascading failures.

Expected Behavior

1. Never write ANY code without requirements + test cases + approval — infra, application, scripts, Dockerfiles, all of it
2. Never give raw commands — always a script with logging, SSO check, error handling
3. Research before implementing — check docs, package managers, auth formats BEFORE writing code
4. Default to requirements — don't ask "should I write requirements?" — just do it
5. Validate assumptions — if code depends on an external system, verify it works on that system BEFORE writing

Acceptance Criteria

• Zero process violations in next session
• Every infrastructure change has requirements → tests → approval → implementation
• Every action delivered as a script, never raw commands
• Research/verification done BEFORE code, not after failure

Cost

Conservative estimate: 6+ hours of user time wasted across Aug 14–20 debugging issues that should never have existed.