Nasiko's agent runtime cuts input tokens 42% by compressing tool outputs
rohanpaul_ai · x · 2026-11-17
Nasiko launched its "Open Runtime for Agents" platform focused on TokenOps and runtime controls. Published benchmarks on gpt-4o over 50 real system-log questions show compression cuts input tokens 42% (617,202→357,958) and cost per correct answer 40% ($0.037→$0.022) with accuracy unchanged (44/50), though a record-lookup task dropped from 100/102 to 96/102. Features include an Agent Gateway, budgets with auto-stop alerts, runtime controls to allow/ask/block every tool call, and an artifact registry.
Related event: Nasiko Launches Open Runtime for AI Agents(5 posts)→
More from coding & agent
- VC Martin Tobias: yesterday's deliverable was a PDF, tomorrow's should be an MCP — MartinGTobias · 2026-11-17
- MIT CSAIL's 0.016% weight-space edit lifts Qwen3.5-0.8B MATH-500 6x without retraining — RandalSchwartz · 2026-11-17
- Neo4j dev advocate to walk Miami JUG through AI agents and MCP — mkheck · 2026-11-17
- Customer saves $9M+ by ditching API pricing and Devin for aggregated AI subscriptions — garrytan · 2026-11-17
- Miro's year-long experiment: making the canvas itself the AI prompt, not a chat sidebar — max__drake · 2026-11-17
- Workato CISO shares how AI agents triage ~80% of security alerts in production — adamse · 2026-11-17