GPT-5.6 Sol always choosing “maximum defense” over a “minimum complete contract”

I’ve been noticing this pattern with GPT-5.6 Sol, both during planning and implementation (today it was on Medium effort). In the first case, the agent implemented a HasOnlyKnownKeys for an external provider response that actually only needed a HasRequiredKeys . That made the consumer reject harmless additional fields (undocumented by the provider) and turned an OK response into an integration failure. In the second case, the agent added many WHERE clauses that overdefined an SQL query. Later, changing something that should have been unrelated created conflicts because the query had encoded redundant assumptions. (curiously it reminded me of something I've learned about technical drawing: adding dimensions that can already be inferred can make the drawing inconsistent when one value changes) I think the common pattern is “maximum defense” instead of a “minimum complete contract”. adding every safeguard that seems reasonable, instead of defining the smallest independent set of rules that fully protects the behavior. I’ve now added global and project-specific instructions to my AGENTS.md files, and created a planning-review skill to challenge redundant inputs, derivable constraints, duplicate sources of truth, and speculative safeguards without making security or correctness loose. did you noticed this with GPT-5.6 Sol or other models? or maybe other cases of agent overengineering that have caused rework and extra token consumption