ck:context-engineering

To optimize LLM token usage, you can analyze context utilization to detect degradation and poisoning risks, then apply compaction and isolation strategies to preserve high-signal context and maintain stable performance.

Context poisoning occurs when irrelevant or corrupted data degrades LLM performance. You can detect it by monitoring context health indicators and evaluating utilization risks across system messages, tools, and history.

For multi-agent coordination, you can partition context across sub-agents using isolation-first workflows and trigger compaction near utilization thresholds to ensure each agent maintains stable context windows.

You can validate context compression quality by using probe-based metrics, including recall, artifact, continuation, and decision testing, to estimate whether the compressed context preserves key outcomes.

You should trigger context compaction in long-running agent workflows when token utilization approaches predefined thresholds, extracting high-signal context to prevent lost-in-the-middle failures before they occur.

Yes, this approach applies to memory systems by evaluating context utilization and applying measurement-before-optimization principles to ensure stable performance across long-running agent workflows and multi-agent coordination.