Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Hermes Agent might have just killed OpenClaw
TL;DR
Alex Finn thinks Hermes is now more reliable than OpenClaw — his core complaint is that OpenClaw’s frequent updates “break” the app and force 30-minute repair sessions, while Hermes ships fewer, more focused releases that fit a clear product narrative.
The standout feature is Hermes’ native Kanban board for true multitasking — instead of one chat thread queuing everything, Alex runs 10–30 parallel task streams across columns like triage, ready, in progress, blocked, and done.
His favorite workflow uses a second “librarian” agent to auto-flesh tasks from memory — every 10 minutes a cron job checks triage, pulls context from his Obsidian vault, expands task details, and moves cards into ready for assignment.
Slashgoal is framed as one of the most underrated AI developments this year — Alex compares it to giving the agent a mission instead of a prompt, with long-running, multi-step tasks that can last minutes, hours, or even 3+ days.
Hermes’ multi-agent profiles are pitched as the fix for memory bloat — Alex recommends at least four agents: a main orchestrator, coding agent, research agent, and administrative assistant, each with separate memories, skills, and environment variables.
A few quieter system features matter a lot for performance — Alex calls out model catalog for brain-and-muscle routing, recommends setting compression threshold to 0.5 to reduce “violent” memory compression, and likes curator pruning unused skills every 7 days.
The Breakdown
Why Alex is jumping ship from OpenClaw
Alex opens with the blunt take: Hermes Agent is currently “a better, more reliable AI agent than OpenClaw.” His frustration with OpenClaw is very specific — constant updates that keep breaking things, plus growing app bloat that slows performance over time. He says the worst part is not even missing features, it’s losing trust: if every update costs him another half hour of fixing, he’d rather stay on an old version.
Hermes’ focused release style wins him over
What he clearly respects about Hermes is discipline. Instead of shipping “a hundred things crammed into one update,” the team pushes smaller, themed updates with a clear narrative, and Alex gives a genuine shoutout to their aesthetics, marketing, and product focus. The vibe here is that reliability is becoming a product feature in itself.
The Kanban board turns Hermes into a real multitasking system
The biggest feature for him is the new Kanban board inside the dashboard, which he sees as the answer to the single-thread bottleneck of Telegram, iMessage, or Discord-based agent chats. He walks through his daily setup: dumping tasks into triage, then letting the system move them through todo, ready, in progress, blocked, and done. For someone juggling SaaS work, content, hiring, and community, he says this is the best native multitasking workflow he’s seen in an agent product.
His custom “librarian” workflow is the real magic trick
Alex’s most practical demo is a second Hermes profile called “librarian” that runs a cron job every 10 minutes. That agent checks triage, pulls relevant context from his Obsidian memory system, fleshes out task details automatically, and moves tasks into ready so his main agent can execute them. He loves that his only job is to create tasks and assign them — the admin work, comments, specs, and output routing happen automatically.
Slashgoal: stop prompting, start assigning missions
From there he shifts to slashgoal, which he calls one of the most underrated AI developments of the year, alongside OpenAI’s Codex. His framing is memorable: a normal prompt asks for actions, while slashgoal gives the agent a mission it can pursue over a long period with many steps and self-testing along the way, “basically like a Ralph loop for your AI agent.” He stresses that prompt quality matters a lot here, and recommends using another LLM to generate a detailed slashgoal prompt first.
Multi-agent profiles keep the system from getting dumb and bloated
Alex then highlights Hermes’ profile system, saying it makes multi-agent setups much easier than OpenClaw. Each profile gets its own memories, skills, and environment variables, which matters because a single agent doing coding, writing, and research will accumulate bloated memory and lose performance. His practical stack is simple: a main orchestrator, a coding agent, a research agent, and an admin agent like librarian.
Small system features that make Hermes feel more mature
He closes with three under-the-radar features he thinks actually matter. Model catalog makes it easy to do his preferred “brain and muscle” setup by assigning different models to different task types and tracking costs; compression can be improved by lowering the threshold to 0.5 so memory compressions happen more often but less aggressively; and curator automatically reviews used vs. unused skills every 7 days and prunes bloat. That last one, he suggests, is exactly the kind of quiet maintenance work OpenClaw could use more of.
The bigger point: Hermes feels dependable right now
By the end, Alex’s thesis is less about flashy capability than about experience. Hermes feels like a tool that’s becoming more structured, more maintainable, and easier to trust day-to-day, while OpenClaw feels stuck in a cycle of fast shipping and breakage. He leaves with an open-ended invitation for more agent content, but the verdict of this video is already pretty clear.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
