LIVE: Hermes vs OpenClaw live tests
TL;DR
Alex Finn’s headline verdict: OpenClaw is still the safer default, but Hermes is the more interesting tinkerer’s tool — after using Hermes for a few weeks instead of “five minutes and a YouTube hot take,” he says OpenClaw works better out of the box, while Hermes shines if you want to customize, inspect tool calls, and push local/open-model workflows.
Hermes’ killer feature is visible self-improvement — Finn loves that Hermes shows every tool call, memory update, file read, and skill creation in detail, with built-in security checks on skill updates, making it far more transparent than OpenClaw.
Hermes’ biggest weakness is memory compaction and self-editing — in Finn’s tests, Hermes repeatedly “lobotomized itself” during compaction, forgot tasks mid-run, and struggled to edit its own files like agent.md without repeated manual intervention and documentation links.
The best workflow may be two agents, not one — instead of replacing OpenClaw, Finn recommends running Hermes alongside it on the same machine, with both agents able to work in parallel and eventually share memory through an Obsidian-based system.
Finn thinks Anthropic is holding back a stronger model—possibly ‘Mythos’—and already using it internally — he argues Claude ships too many strong features too quickly for a normal team using public Opus alone, and predicts the next model tier will be powerful, expensive, and widen the gap between people who can afford frontier AI and those who can’t.
His practical advice is blunt: learn the tools now and make money before elite AI gets pricier — with Claude limits feeling tighter, Opus seeming “stupider” in recent days, and local models becoming more important, he urges viewers to master OpenClaw, Claude Code, Hermes, and local setups before the cost curve leaves people behind.
The Breakdown
The stream opens in full Alex Finn mode: hype, jokes, and a warning against instant hot takes
Finn kicks things off by promising a tour of “the future of AI agents,” comparing Hermes, OpenClaw, and the new orchestration layer Paperclip, with the usual mix of chaos, jokes, and self-aware live-stream energy. But under the bit is a real point: he waited weeks to comment on Hermes because too many creators use a tool for five minutes, then rush out a “greatest ever” or “dead on arrival” video for clicks.
Why Hermes has people excited: transparency, self-improving skills, and local-model friendliness
The first serious case for Hermes is visibility. Finn says Hermes shows every tool call, memory read, memory update, and file access in detail, which makes it dramatically easier to audit than OpenClaw. He’s especially impressed by its self-improving “skills” system: after tasks, Hermes can create or revise skills on the fly and run security checks on those updates, making it feel built for people who want an agent that adapts over time.
A wider detour: Claude, Manus, Perplexity, and the “Anthropic is holding back AGI” theory
Finn zooms out and argues that products like Manus, Perplexity Computer, and Claude’s desktop/computer-use tools are all running in a brutally crowded race—one frontier labs are likely to win. He insists OpenClaw and Hermes remain in a different category because they’re open source and customizable, then launches into his most provocative claim: Anthropic must be using a far stronger internal model—maybe “Mythos,” maybe effectively AGI—to ship Claude features at the pace they have.
The local-model and pricing argument gets more urgent
Another thread running through the stream is cost. Finn says local models are increasingly important, and he entertains a hypothesis from Ray Fernando that energy and hardware costs could push AI subscriptions from $200/month toward $400/month, making local compute more attractive. At the same time, he says Claude’s limits already feel worse than they did months ago, and for the first time he believes the old “models get stupider before the next release” conspiracy might actually be real.
Hermes in practice: fast, durable, and clearly made for tinkerers
Back on Hermes, Finn gives it real credit for performance. Running Opus 4.6 on both tools, he says OpenClaw tends to slow down over time while Hermes has stayed fast, and he likes that Hermes is explicitly designed for people who want to tinker, including with local models like Qwen 3.5. He also contrasts Hermes’ “we trust the user” posture with OpenClaw creator Peter Steinberger’s more skeptical comments about running open models.
The part that frustrated him: compactions were awful and Hermes didn’t understand itself
Then comes the real downside. Finn says Hermes’ compactions were “terrible”: if a compaction happened mid-task, the agent would basically wake up like it had amnesia—“Where am I? What was I doing?”—and derail the whole workflow. Even worse, when he tried to get Hermes to edit its own files, memories, and architecture, it kept making nonsensical changes, like creating a brand-new agent.md in a random directory, forcing him to manually build a custom Obsidian-style memory vault just to stabilize it.
Final recommendation: don’t replace OpenClaw—pair it with Hermes
That leads to his conclusion. If you’re a normal user who doesn’t want to dig through agent.md, memory files, docs, and architecture details, he says he can’t cleanly recommend Hermes over OpenClaw. But if you’re a tinkerer, Hermes is compelling—and for most people, the real win is running both side by side, with Finn already building an Obsidian memory system so Hermes and OpenClaw can share context and work together.
The closing thesis: AI literacy is becoming an economic survival skill
In the last stretch, Finn ties the tool talk to a bigger social argument. If Anthropic really launches a model above Opus that’s too expensive for most people, he thinks the wealth gap widens fast: rich users get the best AI leverage, everyone else falls behind. His answer is education and speed—learn OpenClaw, Claude Code, Hermes, and local workflows now, make money ethically, and get ahead before frontier AI becomes something only a small class can afford.