WORK STORY / 02
Let AI gather evidence before people decide
Context
Information in Notion, Linear, Slack, Swagger, the codebase, and the actual runtime might be unsynchronized or contradictory. A manager’s latest instruction in Slack might not yet appear in formal documentation and might affect existing specification principles. If AI treats every source as equally authoritative, it can easily mistake temporary statements, legacy implementations, or outdated information for formal decisions.
My Decision & Action
I defined different sources as different kinds of engineering evidence: specifications, scope, communication context, API contracts, current implementation, and actual behavior each answered different questions. The Agent first gathered evidence, compared information gaps, and identified conflicts before the relevant people confirmed the actual decision. Important assumptions were then verified through API responses, Playwright, or repeatable measurements.
Outcome
AI no longer modified code immediately after scanning the repository or receiving a single instruction. It first presented evidence, contradictions, and possible impact, prompting the relevant roles to align specifications and update formal sources before deciding where changes were actually needed. This reduced the risk of fixing the wrong layer or allowing historical workarounds to keep spreading.
Reflection
MCP solves information access; Human-in-the-loop solves information adjudication. Tools provide facts, AI assists with organization and inference, people retain final decision authority, and the runtime ultimately verifies whether changes behave as expected.