AI Agents Are Becoming Infrastructure: Paseo, Ruflo, Hermes & OpenClaw

Summary

The video frames AI agents as evolving from chat-based coding assistants into long-running systems that need infrastructure: control planes, persistent memory, skills, communication channels, sandboxes, logs, and recovery mechanisms. It compares four projects by the role each plays in that emerging stack.

Paseo — Remote Control for Coding Agents

Paseo keeps agents and repositories on the user’s machine, then lets users manage them from a phone, desktop, web interface, or CLI. Its local daemon and CLI can launch, list, attach to, message, stop, and inspect agents, with support for isolated worktrees and structured JSON output.

It works with several existing agent tools but does not provide the underlying agents itself. Its security documentation emphasizes localhost defaults, password protection for network access, and caution when exposing the daemon remotely.

Ruflo — Multi-Agent Orchestration

Ruflo is presented as a way to coordinate Claude Code-style agents through specialized agents, swarms, memory, hooks, plugins, background workers, and MCP tooling. It has both a lighter plugin setup and a fuller initialization path.

The video treats its cross-machine federation and security capabilities as project claims that should be verified in practice. A more concrete example cited is a release focused on clarifying overlapping MCP tool descriptions so models choose Ruflo tools appropriately.

Hermes Agent — Memory and Skills

Hermes emphasizes persistent, bounded memory, conversation search, and skills that can be created and improved through use. It supports multiple model providers, messaging gateways, and terminal backends. Native Windows is not supported; the project recommends WSL 2.

The video also notes a release adding scheduled skill-library maintenance, changes to the self-improvement loop, and additional inference-provider support. Hermes can migrate settings and other user data from an OpenClaw directory.

OpenClaw — Personal-Agent Gateway

OpenClaw provides a local-first gateway connecting an assistant to many chat services, tools, voice features, dashboards, and companion apps.

Its broad access makes its security boundary especially important: host tools have full access for the main session by default, while sandboxing can be configured for other sessions. The documentation warns that one shared gateway is not a hostile multi-tenant boundary and recommends separating gateways, credentials, users, or hosts when mixing trust levels. It also recommends auditing exposure, permissions, secrets, plugins, approvals, logging, and sandbox settings.

Safety and Practical Comparison

The video cites a CIK-Bench evaluation of a live OpenClaw instance covering 12 attack scenarios and 88 cases across four backbone models. It reports a 24.6% average baseline attack-success rate, rising to 64–74% when a persistence-state dimension was poisoned.

The video stresses that this is a focused study of persistent-state poisoning, not a general leaderboard or a first-party benchmark comparing all four projects.

The projects’ frequent releases and adoption signals indicate momentum, but not maturity. The suggested fit is:

• Paseo for remote control of existing coding agents.
• Ruflo for experimenting with agent coordination.
• Hermes for persistent memory and skills.
• OpenClaw for a broad personal-agent gateway, after carefully reviewing its security guidance.

Across all four, the video highlights logs, approvals, sandboxing, recovery, and clear ownership of state as essential.

Overall takeaway: The next generation of agent tools will need to make autonomy inspectable, interruptible, and recoverable—not merely offer better prompts or more model choices. When evaluating an agent, consider what it remembers, what it can access, who can reach it, and how to undo a problematic run.

Main Speaker and Sources

• Main speaker: Fluid Coding & AI narrator.
• Sources discussed: The projects’ documentation, READMEs, and release information; the CIK-Bench study on OpenClaw persistent-state safety.