AI & Automation Briefing - August 22, 2026
Nvidia Research: Agent Harness Design Matters More Than Model Choice
Nvidia published research showing that the software harness wrapped around an AI model is more important than the model itself for long-horizon tasks. Researchers paired Claude Opus 5 with a custom harness featuring optimized memory management and a supervisor component, pushing the model's score on the ARC-AGI-3 interactive reasoning benchmark from 30% to 100%. Without the harness, Opus 5 still led all tested models at 30%, but the jump to perfect performance came entirely from harness design, not a model upgrade. A harness is the scaffolding that gives a raw model its tools, memory, context handling, and feedback loops, which is what converts it into a functioning agent. For operations builders designing agentic workflows, this finding shifts the architecture priority: how you build around a model determines performance more than which model you select, particularly for multi-step, multi-day tasks.
FlowEvo Turns Successful Agent Workflows Into Reusable Skills Without Additional Training
A new research paper introduces FlowEvo, an agent architecture that solves a common problem in workflow automation: useful sequences discovered during task execution are normally discarded once the task ends. FlowEvo captures verified successful workflows and converts them into persistent, executable skills that future tasks can reuse or build on. Skills that produce harmful outcomes are identified and suppressed. Running on GPT-4o mini, FlowEvo outperformed eight baseline systems across five benchmarks. On the ALFWorld benchmark it reached an 85.6% success rate, 26.4 percentage points above the next best baseline, while using roughly one third of the tokens. The system was validated across ten different base models, showing that continuous self-improvement is achievable without retraining. For operators building n8n-style pipelines, this research previews an architectural pattern where automation systems accumulate and refine reusable workflow components over time.
OneCLI Launches Open-Source Sandboxed Agent Harness for Teams
OneCLI, a YC S26-backed project, has released an open-source platform that provisions individual sandboxed AI agents for every employee on a team. Each agent operates through a central gateway that injects credentials and enforces team-level policies, keeping access scoped and auditable. The platform supports human-in-the-loop approval workflows and integrates with tools including GitHub, Gmail, Notion, and Dropbox. It is available as a cloud-hosted service at onecli.sh or as a self-hosted deployment. For operations specialists building agent-based workflows, OneCLI is worth evaluating both as a client deployment option and as a reference architecture for credential management and policy enforcement in multi-agent systems.