How to handle an AI agent that outgrows its mandate—and why it’s an opportunity for Indian IT
There is good reason to keep governance outside the loop. Research into ‘context compaction’ (the process by which agents summarize their histories to stay within token-use limits) tested 1,323 episodes across seven model families: violations rose from zero, with constraints in full context, to 30% after compaction, even reaching 59% for some models. If a constraint vanishes in summarization, violations follow. Rules that live only within a conversation can get eaten by it.
