Compaction Cuts Token Use in Automated Malware Analysis Without Changing Overall Scores
Summary
SentinelLABS tested OpenAI’s compaction in automated malware-analysis workflows. It reduced input tokens by about 86% with no measurable change to aggregate scores, though domain object modeling declined.
Key points
- The evaluation compared long-running malware-analysis agent runs with and without context compaction.
- Compaction reduced input tokens by about 86%, output tokens by 31%, and reasoning tokens by 33%; runs used one fewer model call.
- Aggregate evaluation scores were effectively unchanged, but domain object modeling declined.
- The workflow uses compaction for working memory while keeping exact findings, tool outputs, and other artifacts in durable storage.
- Server-side compaction suits automatic handling of long sessions; standalone compaction offers more control at workflow phase boundaries.
- The article recommends evaluating both resource use and task quality, preserving failed paths, and treating compaction as lossy until tested.
Article Details
- Topic
- Evaluating context compaction for automated malware-analysis agents
Vendors
Anthropicpost describing extensions to the Responses API. However, the underlying idea is not unique to OpenAI. Anthropic, Google, and other agent frameworks such as LangChain all expose or document related approaches underOpenAISentinelLABS evaluated OpenAI’s native Responses API implementation against our automated malware analysis evaluation harness to measure real-world impact on task quality and cost.
Products
Codexsummarization and state-carrying systems themselves. The company noted that compaction is the mechanism Codex relies on for long-running coding tasks, which positions it as load-bearing infrastructure rather than aResponses APISentinelLABS evaluated OpenAI’s native Responses API implementation against our automated malware analysis evaluation harness to measure real-world impact on task quality and cost.