Grounding: this subtopic is a synthesis, not a new set of videos. Every claim here is sourced to the §1.1–1.3 articles or to a cross-listed video from Course 1: Picking Your Agent Harness, Course 3: Hermes Agent, or Course 4: Claude Code & AI Coding. Where the synthesis extrapolates beyond the source videos, the article flags it with a > NOTE:.

The fourth subtopic is the one the rest of the course is building toward: when does each harness actually win? The answer is not "the most popular one" or "the cheapest one" — it's a four-way split that depends on your task class, your tolerance for config, your hosting budget, and your need for a 24/7 agent. The matrix below is the cleanest way to read the channel's coverage; the cross-references point you back to the video where each row of the matrix comes from.

What you'll learn

  • The four harnesses split cleanly along four axes: install friction, multi-agent topology, cost ceiling, and 24/7 agent capability. The matrix at the bottom of this article names the winner for each cell.
  • Mavis wins on install friction. Claude Code wins on cost ceiling (when you route through a cheap model). Hermes wins on multi-agent topology. OpenClaw wins on 24/7 agent capability.
  • The "hosted vs self-hosted" axis matters more than the model layer for most users. A hosted variant (MaxClaw, Mavis desktop, the Hermes VPS-Tailscale bridge) gives you a sandbox; a self-hosted variant gives you full control. The channel's framing: start hosted, graduate to self-hosted only when you have a specific reason.
  • The migration path from OpenClaw to Hermes is now a documented operation — the "Migrating from OpenClaw" GitHub script moves soul, memories, and settings. The channel's own team ran it against their Stark agent stack and it worked. This is the path the channel itself is on.
  • The MaxClaw video (42,714 views) is the most-viewed on the channel and the clearest statement of the hosted OpenClaw case. The Mavis video (30,626 views) is the second most-viewed harness video and the clearest statement of the "AI company on easy mode" case.

The four axes

Axis 1 — Install friction

The four harnesses do not require the same effort to get running. From least friction to most:

  1. Mavis — desktop install, one click, $10/month bundled token plan. The channel's framing: "they already set it up for you." (See Article 1.1.)
  2. Claude Code — npm install, authenticate against your Anthropic account, you're done. No daemon, no chat-platform wiring, no model key to pick. (See Article 1.3.)
  3. MaxClaw (hosted OpenClaw) — hosted sign-in with Google at agent.minia.io, pre-configured agent, $20/month. The channel's #1 most-viewed video at 42,714 views. (See Article 1.6 and Course 1 §1.1 — same material.)
  4. Hermes — VPS or local install, the Tailscale bridge for the desktop app, hermes update first, then hermes gateway start, then hermes profile create <role> for every Kanban assignee. (See Article 1.2 and Course 3 §3.2.)
  5. Self-hosted OpenClaw — $2/month Zebra VPS + Mac/Linux install command is the cheap path, but you do the security, the chat-app wiring, and the skill setup yourself. (See Article 1.6 and Course 1 §1.2.)

NOTE: install friction is not the same as ongoing maintenance. Claude Code has the lowest install friction but the highest operational friction (no 24/7 loop, no cron, no Discord bridge). Mavis has low install and low operational friction. Hermes has high install friction but ships the cron + Dashboard UI that lowers operational friction.

Axis 2 — Multi-agent topology

The four harnesses support different multi-agent patterns. From least to most:

  1. Claude Code — single agent. Sub-agents via the --agent-teams flag, but no persistent Kanban, no per-role profiles, no Kanban board.
  2. Mavis — orchestrator + adversarial verifier. Two-worker pattern: a leader agent (the Mavis Assistant) talks to you, a separate verifier audits from first principles with no shared conversation history.
  3. OpenClaw — sub-agents as role-names on top of one model. The channel called out the "Stark deployed a subagent" claim as unfalsifiable until each subagent has a named persona in the run log.
  4. Hermes — Kanban with named profiles per board. Every assignee is a persistent agent under ~/.hermes/profiles/<name>/ with its own SOUL.md, config, and inference provider. The Kanban's parent-child retry loop is the killer feature.

NOTE: Mavis and Hermes share the multi-agent idea but at different points on the curve. Mavis ships a curated two-worker pattern; Hermes is a builder's tool that supports N workers, M boards, and per-profile model choices. The "5x orchestration" claim for Mavis is a cost-shape multiplier on the bundled token plan, not a five-way fan-out.

Axis 3 — Cost ceiling

The four harnesses have very different cost structures. From cheapest to most expensive on the same workload:

  1. Hermes (BYOK free tiers) — free harness, free model tiers (Minimax, Z.AI, Xiaomi Mimo all have free tiers), you pay only what you use. Floor: near-zero for a low-volume daily briefing.
  2. OpenClaw (Minimax coding plan) — $10/month coding plan key, 300 prompts every 5 hours. The channel's recommended cheap path for a 24/7 agent. (See Article 1.6 and Course 1 §1.7.)
  3. Mavis — $10/month bundled token plan, text/image/video under one wallet.
  4. MaxClaw (hosted OpenClaw) — $20/month, MiniMax-backed, pre-configured.
  5. Claude Code on Minimax 2.7 — $10/month coding plan routed through Claude Code's settings.json swap. (See Course 4 §4.2.)
  6. Claude Code on Claude Sonnet/Opus — $5/$25 per million tokens on Opus, $3/$15 on Sonnet, with a Max plan at $200/month.
  7. Perplexity Computer — $176/month on the Max plan, Pro tier blocked from the new product.

NOTE: cost ceiling is not the same as cost per task. A $10/month Mavis plan that gets you through 50 multi-agent trips is cheaper than a $200/month Claude Max plan that you only use for one task a day. The channel's framing: pick the harness by task volume, not by model hype.

Axis 4 — 24/7 agent capability

The four harnesses support different 24/7 patterns:

  1. OpenClaw — yes, with the cron + skills architecture, Discord/Telegram bridge, and the 2-week OpenClaw 3.7 → 4.1 update cadence. The channel's "AI Agent Team that runs your business for $3/month" project lives here.
  2. Hermes — yes, via the Dashboard's cron jobs tab (the most-used surface per the channel) and the Kanban + cron pairing. The Kanban's parent-child retry loop is what makes overnight pipelines reliable.
  3. Mavis — partial. The desktop app is local, the orchestrator runs in the Mavis runtime, but the host's framing is "desktop-first" — no 24/7 daemon story in the public videos yet.
  4. Claude Code — no. It's a CLI, not a daemon. You can run it on a VPS, but you don't get cron, Discord, or daily-briefing out of the box.

NOTE: the Mavis 24/7 story may be evolving — the v0.12 "Curator" release in Hermes is the kind of work that would close the gap. The channel hasn't covered Mavis 24/7 explicitly at the time of writing.

The matrix

A four-by-four grid of "which harness wins" cells. Each cell names the harness that the channel's coverage implies is the right default; the footnote points to the video.

Task class Best harness Why Footnote
I just want a team of agents on my Mac, non-technical Mavis Bundled token plan, curated presets, 5x orchestration §1.1
I want to ship code today, no agent needed Claude Code Low install friction, fast first output, plan mode + agent teams for parallel builds §1.3
I want a 24/7 agent that handles email, files, daily briefings OpenClaw Skills, cron, Discord bridge, the "AI Agent Team that runs your business" pattern §1.6
I want a team of agents on a Kanban Hermes Per-profile roles, parent-child retry, Dashboard cron tab §1.2
I want a sandboxed agent, no install MaxClaw (hosted OpenClaw) Hosted, pre-configured, $20/month, sandboxed §1.6, 42,714-view video
I want to BYOK with free tiers Hermes Free harness + free model tiers (Minimax, Z.AI, Xiaomi Mimo) §1.2
I want to route Claude Code through a cheap model Claude Code + Minimax 2.7 settings.json swap, $10/month coding plan Course 4 §4.2
I want to keep OpenClaw but move hosts OpenClaw + migration script The cross-server migration guide §1.5
I want to migrate from OpenClaw to a successor Hermes (via the "Migrating from OpenClaw" script) Soul, memories, settings transfer §1.5
I want a desktop multi-agent product on Windows Mavis Windows + Mac, one-click install, $10/month §1.1
I want audit surface on every coding task Claude Code inside OpenClaw 3.7/3.8 ACP provenance, session ID, origin tracing Course 4 §4.3
I want to spend $0 on a long-running agent Hermes (BYOK free tiers) or OpenClaw on Minimax $20/month Floor is near-zero on Hermes; floor is $20/month on OpenClaw §1.2, §1.6

NOTE: the matrix is a snapshot, not a permanent ranking. The harness landscape moves every 2 weeks on the OpenClaw side, every release on the Hermes side, and every model price change on the Mavis side. Re-evaluate after 30 days, not 30 minutes.

The hosted vs self-hosted axis

The most under-discussed axis is the one that matters most for new users: hosted vs self-hosted. The four harnesses split as:

  • Hosted only: Mavis (MiniMax-hosted desktop). No self-hosted install path documented.
  • Hosted + self-hosted: OpenClaw (MaxClaw is hosted; the openclaw.ai install is self-hosted). Hermes (the VPS path is self-hosted; the Dashboard is a remote web UI).
  • Self-hosted only: Claude Code (it's a CLI, you run it on your own machine; "hosted Claude Code" is just the Anthropic API).

The channel's framing: start hosted, graduate to self-hosted only when you have a specific reason. A hosted variant gives you a sandbox (the "3am message to my girlfriend" Mac Mini incident in Course 1 §1.1 is the cautionary tale). A self-hosted variant gives you full control. The migration cost from hosted to self-hosted is real — different config files, different skill paths, different update cadence — but the migration cost from self-hosted to hosted is even higher (you lose your custom skills, your custom SOUL.md, and the agent personality you've tuned).

NOTE: the hosted-vs-self-hosted choice is also a security choice. The NemoClaw video (qEFaeLlfLmk, 31,868 views) is the channel's case for a "privacy layer" on top of OpenClaw — the substrate for users who want self-hosted without losing the safety properties of a sandbox. See Article 1.6 for the deeper read.

How to read the matrix in 2026

The matrix at the bottom of this article is a snapshot — it's accurate as of mid-June 2026, but the harness landscape moves faster than any single matrix can capture. Three rules for using the matrix in practice:

  1. The matrix is for the harness decision, not the model decision. The four harnesses in this course can all run on roughly the same set of models (Minimax M2.5/M2.7, Kimi K2.5/K2.7, Qwen 3.5/3.6/3.7, DeepSeek v4, Anthropic Claude Opus/Sonnet). The model-picking framework is in Course 2: AI Models; the matrix in this article assumes you've made the model call separately.
  2. The matrix is for your task, not the channel's task. The channel's most-viewed videos are MaxClaw (42,714 views), NemoClaw (31,868 views), Mavis (30,626 views), and Perplexity vs. Claude Code (18,483 views) — but the channel's primary use case is a multi-agent business rig, not yours. Your use case may be different. Re-read the matrix with your task class in mind.
  3. The matrix is a starting point, not a final answer. The channel's framing: "Re-evaluate after 30 days, not 30 minutes." The four harnesses all have hidden failure modes that show up at week two, not day one. The matrix tells you where to start; the 30-day window tells you where to land.

The two failure modes the matrix doesn't catch

Two failure modes are common enough across all four harnesses that they're worth calling out explicitly, even though they're not in the matrix:

  1. The "I'm not technical" trap. Users who pick a harness based on "I just want it to work" sometimes discover at week two that the harness's curated setup doesn't fit their workflow. The fix is to re-evaluate the harness call, not to fight the harness. The matrix is honest about this — the "best for" column names the use case each harness wins. If your use case doesn't match the column, you picked the wrong harness.
  2. The "I'm too technical" trap. Users who pick a harness based on "I want full control" sometimes discover at week two that the harness's customisation overhead exceeds their time budget. The fix is the same: re-evaluate the harness call. The matrix has a "Builder who wants to own the harness" cell (Hermes), but "Builder who wants to spend their time building" is also a valid persona — and for that persona, Mavis's curated setup is the right pick.

Try it yourself

The hands-on goal for this subtopic: pick three of the four harnesses, run the same task on each, and write the one-sentence harness decision.

  1. Pick a task. Daily news briefing, a single-file coding fix, a research summary, a morning Discord report. Pick one that is small enough to run in under an hour on each harness.
  2. Run it on Claude Code first. Lowest install friction, fastest first output. Note the diff, the time, the cost, and the audit surface.
  3. Run it on Mavis second. One-click install, 5x bundled token plan. Note the orchestrator + verifier pattern, the cost, and the audit surface.
  4. Run it on OpenClaw third. Install via the $2/month Zebra VPS + openclaw.ai install command path from Course 1 §1.2. Note the skills + cron architecture, the Discord bridge, the cost ceiling, and the audit surface.
  5. (Optional) Run it on Hermes. Only if you have time. The Kanban + cron + Dashboard story is the most ambitious harness, but the install friction is real. Skip if you're time-constrained.
  6. Write the one-sentence decision. Format: "I'll use [harness] for [task class] because [reason]." This is the §1.4 capstone sentence.
  7. Commit for 30 days. Re-evaluate after a month of real use, not a weekend of tests. The four harnesses all have hidden failure modes that show up at week two, not day one.

Common pitfalls

  • Picking by view count. MaxClaw is the #1 most-viewed video on the channel (42,714 views), but the right harness for you might be Mavis (30,626 views), Claude Code (18,483 views via the Perplexity Computer test), or Hermes (16,341 views via the Kanban). View count is a popularity signal, not a fit signal.
  • Picking by price. The cheapest harness is not always the right one. A $10/month Mavis plan that gets you through 50 multi-agent trips is cheaper than a $200/month Claude Max plan that you only use for one task a day — but the $200/month plan is the right pick if you need Opus-quality output 8 hours a day.
  • Picking by feature list. Hermes has the most features (Kanban, Dashboard, Skill Bundles, MCP server mode, Desktop App). Mavis has the fewest. Feature list is a depth signal, not a fit signal.
  • Skipping the hosted-vs-self-hosted axis. The 3am message incident in Course 1 §1.1 is the cautionary tale: an unsandboxed agent on a Mac tied to your Apple ID is a real security risk. Start hosted, graduate to self-hosted only when you have a specific reason.
  • Pinning a long-term decision on a single video. The channel's framing on each harness has shifted over time. The Perplexity Computer "KILL" video was followed by a "Claude is back" follow-up. The Mavis 5x claim is a snapshot. The OpenClaw update cadence is 2 weeks. Re-evaluate quarterly.
  • Migrating production agents based on §1.1–1.4 alone. The migration video's own summary: "this video is preliminary testing on the same 'Stark' agent, not a hardened review." Test on a non-production agent first.
  • Ignoring the cost-shape vs cost-ceiling distinction. A flat-rate token plan (Mavis $10/month, OpenClaw Minimax coding plan $20/month) is a cost-ceiling tool; a per-token plan (Anthropic, Minimax pay-as-you-go) is a cost-shape tool. Pick the cost tool that matches your workload.
  • Picking the harness with the best video. The Mavis video is the most-viewed harness video, but the Mavis pattern is not the right pick for every workflow. Read the matrix, not the view count.

Sources

This subtopic synthesises from the §1.1–1.3 articles and the cross-listed videos. The single new video cited:

Plus the cross-references already cited in Article 1.1, Article 1.2, and Article 1.3.