#71 Robin: The End of AI Babysitting - Hermes Agent 0.18, Fable 5, and the Death of Broken Workflows

[SPEAKER_00] You teach an AI your workflow. [SPEAKER_01] Right. [SPEAKER_00] Yeah. [SPEAKER_01] And then you have to fix its mistakes. [SPEAKER_00] Exactly. You fix the mistakes, you come back next week, and you just have to repeat the exact same lesson again. [SPEAKER_01] Which is incredibly frustrating. [SPEAKER_00] It really is. I mean, that is not automation. That is babysitting with extra steps. [SPEAKER_01] It's the core bottleneck of the last three years, honestly. Like, we gave these systems incredible conversational abilities, but we fundamentally failed to give them any reliable executive function. [SPEAKER_00] Well, welcome to this deep dive. I'm really glad you are here with us today. [SPEAKER_01] Yeah. Thanks for joining us. I'm excited for this one. [SPEAKER_00] Today we are exploring a massive update that addresses that exact bottleneck you just mentioned. [SPEAKER_01] We are looking at Hermes Agent 0.18. [SPEAKER_00] Right. Which is powered internally by Fable 5, one of the strongest clawed models available right now. [SPEAKER_01] It really is. [SPEAKER_00] Our mission today is simple. We want to understand how this update shifts AI from a managed tool into a truly autonomous system. [SPEAKER_01] It's a total architectural transformation. The roadmap for today covers how this agent finally learns to think in multitudes. We're going to look at how it permanently remembers your workflows, how it delegates work in parallel, and most importantly, how it rigorously proves its work before calling a task complete. [SPEAKER_00] We are stepping into the era of true agentic workflows. But let's define that up front for clarity. Agendic workflows are AI systems that independently plan and execute multi-step tasks. [SPEAKER_01] Precisely. And achieving that requires fixing the foundational way these models reason. Which brings us to the first major feature. [SPEAKER_00] Right. The mixture of agents protocol. [SPEAKER_01] Exactly. [SPEAKER_00] This feels like an acknowledgement that relying on a single AI model for complex workflows is fundamentally flawed. [SPEAKER_01] It completely is. To do real work, an agent needs more than one perspective. You know, to avoid hallucination blind spots. [SPEAKER_00] Yeah. It's a bit like replacing a single overworked CEO with a diverse board of directors. [SPEAKER_01] That analogy maps perfectly actually. You assign distinct high functioning models to a localized decision council. [SPEAKER_00] So you are utilizing the specialized strengths of different foundation models. [SPEAKER_01] Right. Like, Claude is fantastic at holding long context. GQT brings a structured analytical angle. And DeepSeq catches obscure logic errors the other two might miss. [SPEAKER_00] And you just set this up in the configuration. [SPEAKER_01] Yeah. You create a preset. You might literally name it Council and select your models. From there, you designate an aggregator model to oversee them. [SPEAKER_00] And just to clarify, an aggregator model is a master model that combines different answers into one best response. [SPEAKER_01] Right. And the way it combines them is the actual breakthrough. [SPEAKER_00] How so? [SPEAKER_01] The reference models process the prompt entirely independently and parallel. They formulate their own distinct chains of thought. [SPEAKER_00] I am curious about that reconciliation process though. [SPEAKER_01] Okay. [SPEAKER_00] If the models completely disagree on the core architecture, how does the aggregator handle conflicting advice? [SPEAKER_01] So it weighs the underlying reasoning to find the strongest path. Mm-hmm. It analyzes the logical bridges each model built to reach its conclusion and synthesizes the surviving, logically sound arguments. [SPEAKER_00] So the aggregator filters out the noise for us. [SPEAKER_01] It does. It forces the models to peer review each other's homework. [SPEAKER_00] But even a brilliant council of models is going to fail if they walk into the boardroom with total amnesia. [SPEAKER_01] Ah, the persistent memory problem. [SPEAKER_00] Yeah. I mean, I still wrestle with AI for getting my formatting rules every single time. [SPEAKER_01] It drives everyone crazy. [SPEAKER_00] It really does. You spend 30 minutes establishing a perfect structure only for it to reset the next day. [SPEAKER_01] Well, Hermes 0.18 addresses this directly with a new command called slash learn. Slash learn. Yeah. It's designed to build a compounding skill library. There are two ways it learns. First, you can point it at an authoritative source like an SOP or a code repo to learn house rules. [SPEAKER_00] Okay. That makes sense. [SPEAKER_01] But the second method is the real leap. You can ask it to learn from a multi-step conversation you just finished. [SPEAKER_00] Like researching a tool, checking pricing, and formatting a review page. [SPEAKER_01] Exactly. It turns a good one-time result into a reusable button. [SPEAKER_00] But doesn't this clutter the system? I mean, if you save a chat, does it save all the messy trial and error? [SPEAKER_01] No, not at all. It specifically pulls only the successful sequence of steps from the conversation. [SPEAKER_00] It extracts the winning formula, ignoring the messy chat. [SPEAKER_01] You're literally turning one-off successes into permanent assets. [SPEAKER_00] Which sounds incredible. But compounding systems carry a specific risk. If it's constantly learning, how do we stop it from holding on to an outdated rule? [SPEAKER_01] That's where the companion feature comes in. It's called Slash Journey. [SPEAKER_00] Slash Journey. [SPEAKER_01] Yeah. It opens a visual map of the skills and memories the agent has built. Oh, wow. Memory is only useful when it's accurate, right? If the agent saves an old rule about how review pages should be structured but your format changes, Slash Journey lets you find that memory. [SPEAKER_00] And edit it. [SPEAKER_01] Exactly. Yeah. You edit it in line or delete it. This prevents one wrong assumption from ruining future work. [SPEAKER_00] It's like pruning a bonsai tree so it grows in the right direction, turning self-improvement into something you can actually manage. [SPEAKER_01] That's a phenomenal way to visualize it. [SPEAKER_00] But practically speaking, how do we know when a memory has become outdated and needs checking? [SPEAKER_01] Well, you'll notice it when outputs start drifting, prompting you to check the visual map. [SPEAKER_00] We spot the bad output, then prune the memory tree. [SPEAKER_01] Exactly. It gives you control. [SPEAKER_00] So, we have smart reasoning and a clean memory. But how does it physically execute a massive task without making us wait forever? [SPEAKER_01] That brings us to background delegation or fan out. [SPEAKER_00] Fan out. [SPEAKER_01] Right. Instead of moving in a straight line where the agent finishes part one, moves to part two, and you just watch the dead time. [SPEAKER_00] Which takes forever. [SPEAKER_01] It does. Hermes agent spins up multiple sub-agents in the background instead. [SPEAKER_00] Okay. [SPEAKER_01] So, sub-agent one researches features and pricing. Sub-agent two scans Reddit and X for real user reactions. [SPEAKER_00] Operating in parallel. [SPEAKER_01] Yes. And crucially, sub-agent three waits for both inputs before building their review page. [SPEAKER_00] Whoa. Imagine setting three agents loose simultaneously and just waiting for the finished product. [SPEAKER_01] I know. It's a massive leap in efficiency. [SPEAKER_00] But what physically stops sub-agent three from jumping the gun and writing a half-baked page early? [SPEAKER_01] It has a strict built-in dependency step requiring all inputs first. [SPEAKER_00] It won't bake the cake until all ingredients arrive. [SPEAKER_01] Right. The architecture physically prevents it from starting until the data is fully resolved. [SPEAKER_00] That is fascinating. [SPEAKER_01] Yeah. [SPEAKER_00] But working fast in parallel is great. How do we know the final product actually works before we trust it blindly? [SPEAKER_01] This is the most important upgrade in 0.18 in my opinion. [SPEAKER_00] Okay. [SPEAKER_01] Usually an agent says it's done, but then you test it and the form doesn't submit or the API fails. [SPEAKER_00] Right. You still have to act as quality control. [SPEAKER_01] Exactly. Completion contracts fix this. Instead of a vague goal, you write success checks. [SPEAKER_00] Success checks. [SPEAKER_01] Yeah. A judged model reviews the result against the contract after every turn. If the form still fails, Hermes Agent cannot report the task as finished. It keeps working. [SPEAKER_00] I have to push back a bit here though. Isn't this just a fancy way of saying we need to write better, longer prompts? [SPEAKER_01] It's a fair question, but no. Prompts just give instructions. Completion contracts are a distinct verification layer testing the physical output in a sandbox. [SPEAKER_00] But could an agent just get trapped in an endless loop trying to fix an impossible bug? [SPEAKER_01] There are safeguards, but essentially it works iteratively to fix the issue until the contract criteria are unequivocally met. [SPEAKER_00] It refuses to quit until the proof is undeniable. [SPEAKER_01] Exactly. It forces the AI to mathematically prove its competence. Answer. So, if we pull all this together, Hermes Agent 0.18 represents a real paradigm shift. [SPEAKER_00] It really does. I mean, with a council of models, a clean skill library, parallel sub-agents and completion contracts, we are finally moving from a tool that requires constant supervision to a system you can walk away from. [SPEAKER_01] The ultimate takeaway is just less babysitting, more trust. [SPEAKER_00] Absolutely. Which leaves us with a final thought to mull over. [SPEAKER_01] Yeah. [SPEAKER_00] If AI can now verify its own work, manage its own memory and delegate to its own sub-agents, what exactly does our human role become? [SPEAKER_01] Are we shifting from being the doers to simply being the ones who write the completion contracts of the future? [SPEAKER_00] It is definitely something to think about. Thank you for taking the time to explore this deep dive with us today. We really appreciate you being here. [SPEAKER_01] Yeah. Thanks for listening. Keep exploring.