a funny agentic failure i've found a while back the main agent have several subagents as tools, i prompted the main agent to delegate an intensive multi step task to one of the "specialized" subagents for that task. The subagent kept looping over it's tools until it reached the harness's turn limit, at that point the harness dropped the oldest message from it's context to accomodate the turns limit ( the initial user message from main agent ). The resulting request contained system, assistant, and tool messages, but no user turn (the initial message from main agent), strict input validation at the LLM gateway correctly rejected it with "No user query found in messages" the main agent caught the correctness bug as upstream exception and the whole request to the gateway was dumped into SSE as a system error that included the full system prompts, tools schemas and llm gateway api keys. Agents turn limit and context compaction paths could become security boundaries