Julian Goldie demos paid 'Agent OS' that runs Claude, Hermes, Codex and Antigravity from one dashboard

• Agents named in the video: Claude, Hermes, Codex and Antigravity, all in one hub (00:09). The v2 blog post lists a different set: Claude Opus profiles, "GPT-6 Astra" in a Codex tab, OpenClaw, and free-model routers (blog). No exact model versions are given in the video.
• Modules:
• Search OS for keyword research (01:00)
• Content writing and deployment
• Video creation
• Voice control (03:16)
• "Mastermind", where several agents work on one task (03:34)
• 3D worlds built with Codex and Blender (04:11)
• A pixel-art office view of the agents (04:33)
• Sales-call scoring against a company's SOP (standard operating procedure) (05:00)
• OpenMontage video (05:18)
• Access: members only, as a zip with a 30-day setup roadmap (02:14; Skool).

• What is documented: the v2 post describes four layers: model tabs, an Obsidian memory that agents write to automatically, workflow agents (video built on Remotion, SEO, outreach), and a "builder versus critic" check on outputs (blog). No code, API surface, orchestration protocol or licence is published.
• "Mastermind": described only as agents "talking to each other" (03:43). How messages pass between agents and how conflicts are resolved is not shown.
• Components: the underlying pieces are inspectable on their own terms. OpenMontage drives a coding agent through YAML pipeline manifests, Markdown "director skills" and Python tools. It has self-review checkpoints and approval gates for a human (GitHub).
• Reproducibility: none for the hub as a whole.

• Productivity and quality: claims that an article is written and deployed with one click, or that the video module produces "cinematic" output, come only from the creator's demo (01:26, 05:21). They are sales material, not measurements, and no output samples or benchmarks are published.
• Member numbers: the post claims 3,600+ members (Reddit), while Goldie's site says 3,000+ (blog). Both are self-reported.
• Hermes and OpenMontage traction: reported by press and GitHub. Star counts differ by source and date: 32k in Pinggy versus 71.8k on the repo now.
• No independent coverage or test of the hub itself was found in the reviewed sources.

Putting several agents behind one control panel is already a crowded idea. Antigravity's Manager View orchestrates parallel agents. Hermes positions itself as a persistent personal agent that learns skills, and succeeds OpenClaw. OpenMontage turns any coding agent into a video pipeline. Goldie's "OS" mostly bundles these tools into an interface tuned for SEO-agency work.

• Inconsistent descriptions: the name of the agent-office feature is "Staret" in the post but "Starnet" in the video (04:33). The list of agents also differs between the video and the v2 post.
• Gimmick, by the creator's own admission: Goldie concedes the pixel office is partly a gimmick (04:50). He also says he spends 3–4 hours a day maintaining the system (05:39).
• Licensing: OpenMontage is AGPL-3.0 (GitHub). That matters if it is redistributed inside a paid bundle. The reviewed sources don't say how the zip handles this.
• Security: agents that auto-deploy content and process call recordings raise data-handling questions that none of the sources address.

• Copy the pattern, not the bundle: shared memory, separate tabs per role, and a builder agent checked by a critic agent are all reasonable. Each can be built from documented open tools such as Hermes, OpenMontage and Obsidian without buying an undocumented zip.
• Budget for upkeep: the creator's own 3–4 hours a day is a realistic signal of what maintaining a home-built orchestration layer costs.
• Sales-call scoring against SOPs is the module most useful to a business, but there is no evidence of its accuracy. Test it against human review before relying on it.
• Read it as marketing: the urgency framing (06:23) is a funnel for the community, not evidence that the system works.