LIVE: MASSIVE Claude Code leak fallout. + Hermes agent workflows
TL;DR
The Claude Code leak convinced Alex Finn that Anthropic is building toward autonomous, proactive agents — his big takeaway from the exposed code is a future where Claude doesn’t just build a to-do app, but keeps going on its own with calendars, invites, themes, and integrations.
He’s betting Anthropic launches a $1,000-$2,000 Claude plan within four weeks — Alex argues recent usage-limit squeeze plus leaked references to new models like Mythos/Capybara point to a higher-priced tier for heavier agentic workloads.
Alex thinks the real divide in AI will be economic, not just technical — his claim is that when top-tier models like the rumored Mythos arrive, people with cash to deploy them will compound advantages fast, creating a “permanent underclass” among those who don’t keep up.
Hermes is faster and better at spinning up multiple specialized agents, but OpenClaw still wins for memory reliability — his current setup uses OpenClaw as the orchestrator, Hermes as worker agents, and even a fully local Hermes agent named Ruben running Qwen 3.5 397B on a 512GB Mac Studio.
He’s building a custom multi-agent memory layer in Obsidian — the system uses separate folders for OpenClaw, Hermes, and shared context so agents can both message each other directly and leave durable notes, mistakes, guides, and daily logs for cross-agent coordination.
His broader message is to stop copying crowded AI ideas and chase uncrowded “moonshot” problems instead — rather than building another vibe-coding wrapper or social posting tool, he says this is the rare moment to run a race with zero competitors and massive upside.
The Breakdown
Leakgate Starts With a Buddy Gacha Detour
Alex opens in full live-stream chaos mode: the “massive Claude Code leak,” a promised Hermes deep dive, and repeated hype for what he calls the biggest announcement of his life coming Monday. Before any serious analysis, he gets sidetracked hatching Claude’s new /buddy pet, gets a useless ghost named Squib, and immediately jokes that this might make him “team Codex.”
The Leak’s Biggest Signal: Claude That Keeps Going Without You
Once he gets into the actual leak, Alex says the most important reveal is “proactive mode.” His example is simple but sticky: ask Claude to build a to-do app, and instead of stopping there, it keeps layering on calendars, sharing, dark mode, and integrations on its own. That, to him, is where agents are headed — but he says it only works if Anthropic dramatically raises limits.
Why He Thinks Anthropic Is About to Sell a $2,000 Plan
Alex’s strongest prediction is that Anthropic is deliberately squeezing current Claude plan limits to tee up a much pricier tier. He repeatedly says a $1,000 or $2,000 plan lands within four weeks, bundled with a new frontier model — Mythos, allegedly referenced in the leak as “Capybara.” He’s blunt about it: Claude Code already creates enough economic value that he’d buy the expensive plan “day one.”
New Models, Dream Mode, and the Rich-Get-Richer Warning
He runs through other leaked hints: Sonnet 4.8, Opus 4.7, Mythos/Capybara, and a “Dream Mode” where Claude keeps thinking overnight like an employee burning tokens while you sleep. That leads to one of his bigger social arguments: if the best models become expensive enough, only well-capitalized users get access, and they widen the gap even more. His advice is practical and urgent in his framing — make money and save now so you can afford the next generation of tools.
Hermes vs. OpenClaw: Fast Worker vs. Reliable Brain
The stream then pivots into his live agent stack. Alex says Hermes has rough edges, especially around memory and compaction, but it’s faster and now makes it easy to spin up separate agents with distinct workspaces, tools, and skills. His preferred setup today is OpenClaw as the orchestrator and Hermes as specialized workers, including a fully local Hermes agent named Ruben powered by Qwen 3.5 397B on his Mac Studio.
His Obsidian Memory System Adds a Second Layer of Agent Coordination
Alex previews a custom Obsidian-based memory system he already shared inside Vibe Coding Academy. The structure is straightforward: one folder for OpenClaw, one for Hermes, and one shared folder, each containing things like daily logs, guides, mistakes, and working context. In his view, this gives agents two ways to cooperate — direct communication through ACP plus a persistent shared knowledge layer they can all read from.
Smart Glasses, Local Models, and Why Tinkering Beats Luxury
One of the most fun segments is a live demo of his Even Reality G2 glasses, which he hacked so his main OpenClaw agent, Henry, can execute commands from voice prompts. He tells the glasses to open Reddit on his computer hands-free, then screen-shares the result to prove it worked. The bigger point is personal: Alex says getting local Qwen models working on his desk and talking to agents through glasses made him happier than money ever has.
Don’t Enter Crowded AI Races — Find the Empty One
He closes on a broader founder rant: too many people are building “another” vibe-coding tool, “another” social media tool, or “another” wrapper. His repeated metaphor is racing — why line up next to Usain Bolt when there are giant new races with almost nobody in them? That theme ties back to Monday’s teased reveal, which he frames as his own high-risk swing at a category-defining, paradigm-shifting problem rather than an easy AI single.