AI Employee Slashes Thread Costs by 82% Using Prompt Caching Architecture

Viktor, an AI employee built for Slack and Microsoft Teams, slashes agent thread costs by 81.8% on Claude Opus 4.8 through a prompt caching architecture that drops a 40-step thread from $11.35 to just $2.07 by keeping prefixes byte-stable, treating threads as append-only logs, and timing cache compaction strategically around idle windows before the 5-minute TTL expires.

• Viktor, an AI employee operating in Slack and Microsoft Teams, reduces agent thread costs by 81.8% on Claude Opus 4.8 by architecting its entire thread engine around prompt caching, dropping a 40-step thread from $11.35 to $2.07.
• To keep the prompt prefix byte-stable and cache-friendly, tools are exposed as SDK functions in code rather than injected JSON schemas, and every thread is treated as an append-only log where history is never edited or reordered.
• Compaction runs inside the thread's own warm cache to avoid paying full re-processing costs, and is timed strategically: suppressed during active threads to protect the cache, then triggered aggressively around the 3-minute idle mark before the 5-minute cache TTL expires.