Latest / AI Fire Daily / #71 Robin: The End of AI Babysitting - Hermes Agent 0.18, Fable 5, and the Death of Broken Workflows
Transcript
- 0:00You teach an AI your workflow. Right. Yeah. And
- 0:03then you have to fix its mistakes. Exactly. You
- 0:04fix the mistakes, you come back next week, and
- 0:06you just have to repeat the exact same lesson
- 0:08again. Which is incredibly frustrating. It really
- 0:10is. I mean, that is not automation. That is babysitting
- 0:13with extra steps. It's the core bottleneck of
- 0:16the last three years, honestly. Like, we gave
- 0:18these systems incredible conversational abilities,
- 0:21but we fundamentally failed to give them any
- 0:24reliable executive function. Well, welcome to
- 0:27this deep dive. I'm really glad you are here
- 0:29with us today. Yeah, thanks for joining us. I'm
- 0:31excited for this one. Today, we are exploring
- 0:34a massive update that addresses that exact bottleneck
- 0:37you just mentioned. We are looking at Hermes
- 0:40Agent 0 .18. Right, which is powered internally
- 0:44by Fable 5, one of the strongest Claude models
- 0:47available right now. He really is. Our mission
- 0:50today is simple. We want to understand how this
- 0:52update shifts AI from a managed tool into a truly
- 0:57autonomous system. It's a total architectural
- 1:00transformation. The roadmap for today covers
- 1:02how this agent finally learns to think in multitudes.
- 1:06We're going to look at how it permanently remembers
- 1:08your workflows, how it delegates work in parallel,
- 1:11and most importantly, how it rigorously proves
- 1:14its work before calling a task complete. We are
- 1:17stepping into the era of true agentic workflows.
- 1:21But let's define that up front for clarity. Agendic
- 1:24workflows are AI systems that independently plan
- 1:27and execute multi -step tasks. Precisely. And
- 1:31achieving that requires fixing the foundational
- 1:33way these models reason. Which brings us to the
- 1:36first major feature. Right, the mixture of agents
- 1:38protocol. Exactly. This feels like an acknowledgement
- 1:41that relying on a single AI model for complex
- 1:44workflows is fundamentally flawed. It completely
- 1:47is. To do real work, an agent needs more than
- 1:49one perspective, you know, to avoid hallucination
- 1:51blind spots. Yeah, it's a bit like replacing
- 1:53a single overworked CEO with a diverse board
- 1:56of directors. That analogy maps perfectly, actually.
- 1:59You assign distinct high -functioning models
- 2:01to a localized decision council. So you are utilizing
- 2:03the specialized strengths of different foundation
- 2:06models. Right, like... Claude is fantastic at
- 2:09holding long context. GKT brings a structured
- 2:11analytical angle, and DeepSeq catches obscure
- 2:15logic errors the other two might miss. And you
- 2:17just set this up in the configuration? Yeah.
- 2:19You create a preset. You might literally name
- 2:21it Council and select your models. From there,
- 2:24you designate an aggregator model to oversee
- 2:27them. And just to clarify, an aggregator model
- 2:30is a master model that combines different answers
- 2:32into one best response. Right. And the way it
- 2:36combines them is the actual breakthrough. The
- 2:38reference models process the prompt entirely
- 2:40independently in parallel. They formulate their
- 2:43own distinct chains of thought. I am curious
- 2:46about that reconciliation process, though. If
- 2:49the models completely disagree on the core architecture,
- 2:53how does the aggregator handle conflicting advice?
- 2:55So it weighs the underlying reasoning to find
- 2:58the strongest path. It analyzes the logical bridges
- 3:01each model built to reach its conclusion and
- 3:03synthesizes the surviving logically sound arguments.
- 3:07So the aggregator filters out the noise for us.
- 3:10It does. It forces the models to peer review
- 3:13each other's homework. But even a brilliant council
- 3:15of models is going to fail if they walk into
- 3:18the boardroom with total amnesia. Ah, the persistent
- 3:21memory problem. Yeah. I mean, I still wrestle
- 3:23with AI for getting my formatting rules every
- 3:26single time. It drives everyone crazy. It really
- 3:28does. You spend 30 minutes establishing a perfect
- 3:30structure only for it to reset the next day.
- 3:34Well, Hermes 0 .18 addresses this directly with
- 3:37a new command called slash learn. Yeah, it's
- 3:41designed to build a compounding skill library.
- 3:45There are two ways it learns. First, you can
- 3:47point it at an authoritative source, like an
- 3:49SOP or a code repo, to learn house rules. Okay,
- 3:52that makes sense. But the second method is the
- 3:54real leap. You can ask it to learn from a multi
- 3:57-step conversation you just finished. Like researching
- 3:59a tool, checking pricing, and formatting a review
- 4:02page. Exactly. It turns a good one -time result
- 4:05into a reusable button. But doesn't this clutter
- 4:07the system? I mean, if you save a chat, does
- 4:10it save all the messy trial and error? No, not
- 4:12at all. It specifically pulls only the successful
- 4:15sequence of steps from the conversation. It extracts
- 4:18the winning formula, ignoring the messy chat.
- 4:21You're literally turning one -off successes into
- 4:24permanent assets. Which sounds incredible. But
- 4:28compounding systems carry a specific risk. If
- 4:31it's constantly learning, how do we stop it from
- 4:34holding onto an outdated rule? That's where the
- 4:36companion feature comes in. It's called slash
- 4:38journey. Slash journey. Yeah. It opens a visual
- 4:41map of the skills and memories the agent has
- 4:43built. Oh, wow. Memory is only useful when it's
- 4:46accurate, right? If the agent saves an old rule
- 4:49about how review pages should be structured,
- 4:51but your format changes, slash journey lets you
- 4:55find that memory. And edit it. Exactly. Yeah.
- 4:58You edit it in line or delete it. This prevents
- 5:00one wrong assumption from ruining future work.
- 5:03It's like pruning a bonsai tree so it grows in
- 5:04the right direction, turning self -improvement
- 5:07into something you can actually manage. That's
- 5:09a phenomenal way to visualize it. But practically
- 5:12speaking, how do we know when a memory has become
- 5:14outdated and needs checking? Well, you'll notice
- 5:17it when outputs start drifting, prompting you
- 5:19to check the visual map. We spot the bad output,
- 5:22then prune the memory tree. Exactly. It gives
- 5:24you control. So, we have smart reasoning and
- 5:27a clean memory. But how does it physically execute
- 5:31a massive task without making us wait forever?
- 5:34That brings us to background delegation, or fan
- 5:37out. Fan out. Right. Instead of moving in a straight
- 5:40line where the agent finishes part one, moves
- 5:42to part two, and you just watch the dead time.
- 5:44Which takes forever. It does. Hermes agent spins
- 5:46up multiple subagents in the background instead.
- 5:49OK. So subagent one researches features and pricing.
- 5:53Subagent two scans Reddit and X for real user
- 5:56reactions. Operating in parallel. Yes. And crucially,
- 6:00subagent 3 waits for both inputs before building
- 6:03the review page. Whoa. Imagine setting three
- 6:05agents loose simultaneously and just waiting
- 6:08for the finished product. I know. It's a massive
- 6:10leap in efficiency. But what physically stops
- 6:13subagent 3 from jumping the gun and writing a
- 6:16half -baked page early? It has a strict built
- 6:19-in dependency step requiring all inputs first.
- 6:22It won't bake the cake until all ingredients
- 6:24arrive. Right. The architecture physically prevents
- 6:27it from starting until the data is fully resolved.
- 6:30That is fascinating. But working fast in parallel
- 6:33is great. How do we know the final product actually
- 6:37works before we trust it blindly? This is the
- 6:40most important upgrade in 0 .18 in my opinion.
- 6:43Okay. Usually an agent says it's done, but then
- 6:45you pest it and the form doesn't submit or the
- 6:48API fails. Right. You still have to act as quality
- 6:51control. Exactly. Completion contracts fix this.
- 6:54Instead of a vague goal, you write success checks.
- 6:56Success checks. Yeah. A judge model reviews the
- 6:59result against the contract after every turn.
- 7:02If the form still fails, Hermes' agent cannot
- 7:05report the task as finished. It keeps working.
- 7:09I have to push back a bit here, though. Isn't
- 7:11this just a fancy way of saying we need to write
- 7:13better, longer prompts? It's a fair question,
- 7:15but no, prompts just give instructions. Completion
- 7:18contracts are a distinct verification layer,
- 7:20testing the physical output in a sandbox. But
- 7:23could an agent just get trapped in an endless
- 7:25loop trying to fix an impossible bug? There are
- 7:28safeguards, but... Essentially, it works iteratively
- 7:31to fix the issue until the contract criteria
- 7:34are unequivocally met. It refuses to quit until
- 7:36the proof is undeniable. Exactly. It forces the
- 7:39AI to mathematically prove its competence. Answer.
- 7:43So, if we pull all this together, Hermes agent
- 7:460 .18 represents a real paradigm shift. It really
- 7:51does. I mean, with a council of models, a clean
- 7:54skill library. parallel subagents, and completion
- 7:57contracts, we are finally moving from a tool
- 8:00that requires constant supervision to a system
- 8:02you can walk away from. The ultimate takeaway
- 8:04is just less babysitting, more trust. Absolutely.
- 8:07which leaves us with a final thought to mull
- 8:10over. Yeah. If AI can now verify its own work,
- 8:13manage its own memory, and delegate to its own
- 8:15subagents, what exactly does our human role become?
- 8:19Are we shifting from being the doers to simply
- 8:21being the ones who write the completion contracts
- 8:23of the future? It is definitely something to
- 8:25think about. Thank you for taking the time to
- 8:26explore this deep dive with us today. We really
- 8:28appreciate you being here. Yeah. Thanks for listening.
- 8:31Keep exploring.