Token savings versus task success when trimming AI-agent context

DAIR.AI says a study compared five context-trimming strategies on multi-step tool workflows. Recency, relevance and summarization saved about 60% of tokens, but task success fell to 66.6%–77.3%.

Protocol-aware trimming preserved identifiers, constraints, tool schemas and unresolved commitments while compressing the rest. With adaptive budget guardrails, it saved 56.0% of tokens, achieved 96.0% task success and recorded 1.0% cascading failure.

The summary also highlights the importance of retaining enough context: keeping 25% or less raised the odds of failure 10.92 times compared with keeping 50% or more. Complex workflows needed more retained context. The caveat is that protected state came from gold annotations; a production system would still need to detect that state itself.