{
  "id": 3247194,
  "title": "Passing Once Isn't Reliable — This Week's Agent Engineering Puts the Harness Before the Model",
  "url": "https://urgent.news/2026/08/25/passing-once-isnt-reliable-this-weeks-agent-engineering-puts-the",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-08-25T11:14:24.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/felipe0liveira/passing-once-isnt-reliable-this-weeks-agent-engineering-puts-the-harness-before-the-model-3maj"
  },
  "original_language": "en",
  "account": "This week's AI agent developments focus on orchestration patterns, tool/function calling, memory, planning loops, multi-agent coordination, and agent evaluation. AgentWeave, a new system, filters tool sets before prompts are processed, reducing tool exposure by 70%, input tokens by 62%, and latency by 51%. However, pass@1 success rates in Thinkingbox, a sandbox and benchmark, do not guarantee reliability as seen in the drop from 65.36% pass@1 to 25.25% pass@20. Spine-Branch Coordination for Multi-agent Computer Use tackles VM state merging in multi-agent systems, lifting success rate by 6-16.5 points and cutting cost per task by 34-70%. AutoSaddler optimizes the agent's external harness using execution traces, resulting in gains of 9-10 points across three benchmarks. LangSmith Tuned Evaluators, starting with Perceived Error, grade 100% of production traffic, cutting evaluation cost by 82% while maintaining quality. Anthropic Engineering Blog and LangChain / LangGraph Blog introduced new tools and frameworks for testing and evaluating agents, emphasizing the importance of the agent harness and its impact on performance.",
  "summary": "This digest covers AI agent developments from 2026-08-18 to 2026-08-25: orchestration patterns, tool/function calling, memory, planning loops, multi-agent coordination, and agent evaluation. 🔥 Highlights AgentWeave: Routing Before Reasoning for Efficient Function Calling in Tool-Rich Language Models — cuts tool exposure 70%, latency 51%. One Success Isn't Reliability: Thinkingbox, a Sandbox and…",
  "key_points": [
    "AgentWeave reduces tool exposure by 70% and latency by 51%",
    "Pass@1 in Thinkingbox drops from 65.36% to 25.25% despite initial success",
    "AutoSaddler optimizes agent harness, gains 9-10 points across benchmarks"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}