Firstly, 45% of people deploy Hermes and struggle to even run it once.
The next 45% run it for once, test it, and then forget about it.
The next 5% run it well, without knowing everything (especially the most important things)
If you are typing a prompt, getting an answer, and closing the tab, your operational unit economics are broken.
You are basically treating
a fully-loaded autonomous architecture like a slightly smarter ChatGPT.
When an intern was architecting the data pipelines and UI for BoutPredict (my AI combat sports prediction platform), I thought his bottlenecks would be Snowflake queries and Airflow DAGs.
They weren’t.
His bottleneck was how he was interacting with the AI tooling.
I was manually providing context, repeating instructions, and restarting sessions every time I needed to pivot from Python backend logic to Vue.js frontend scaffolding.
Two developers recently published a brutal technical teardown of how the majority of users are deploying Hermes (a provider-agnostic AI agent that runs across 17 platforms)
Their core thesis:
95% of users have never touched the agent’s actual capabilities.
The persistent memory, the session branching, the file rollbacks, the multi-model routing , it all sits there untouched.
Here is the 12 structural features that separate the tourists from the actual engineers.
If you are not using these, you are the human holding your own projects back.
The Persistent Architecture
1. SOUL.md & /personality
Stop typing “You are a senior Vue.js developer” at the start of every chat. Hermes reads SOUL.md at boot. Whatever is in that file becomes your agent's permanent baseline behavior across every session and platform. You can use /personality to hot-swap named personas on the fly. You write the identity once.
2. MEMORY.md & USER.md
Context collapse is the death of productivity. Hermes utilizes two persistent files read at every session. MEMORY.md acts as the project notebook, and USER.md tracks what the agent knows about your specific coding habits. Because it is indexed with FTS5 and an LLM summarizer, a hyper-specific deployment fix you discovered eight weeks ago automatically surfaces in today's session. You stop re-explaining your architecture.
Execution & Version Control
3. /branch (Session Forking)
You are deep into a massive refactor, and you want to try a risky architectural pivot. Instead of burning your warm context or starting a new chat, you use /branch. It forks the session exactly like a Git commit. If the experiment fails, you seamlessly revert to the main branch.
4. /rollback (Filesystem Safety)
If the agent hallucinates and nukes a critical configuration file, you don’t even need to touch Git. Hermes saves a snapshot of every single file it touches. Running /rollback acts as an instant filesystem checkpoint.
5. Mid-Flight Steering (/steer & /queue)
If your agent is three tool calls deep into a massive run and you realize it is targeting your production database instead of staging, you do not kill the process. You type /steer use staging not prod. The agent intercepts the note on the next tool call without breaking the cache. /queue allows you to line up the subsequent instructions without interrupting the current execution.
6. The Power Toggles (/yolo, /fast, /reasoning)
Most developers stay on default settings and complain about speed. These toggles alter the agent’s core operational parameters. /yolo bypasses dangerous-command human approvals for rapid prototyping. /fast flips the API to OpenAI Priority or Anthropic Fast Mode. /reasoning manually dictates the compute effort for heavy logical tasks.
The Routing Engine
7. Model Swapping on the Fly (/model)
You are not locked into one vendor. You can run /model anthropic:claude-opus-4-7 for a heavy architectural lift, and instantly swap to /model openrouter:kimi-k2.6 for cheap, bulk data processing—all within the exact same session, without losing state.
8. Auxiliary Model Routing
Agents perform massive amounts of background work: compressing context, summarizing sessions, generating chat titles. The developer noted that if you route all of this through a frontier model, your API costs will bleed you dry. Hermes allows you to route the main logic to Opus 4.7, while offloading background compression to a much cheaper model like Haiku.
The Distribution & Automation Layer
9. The 17-Platform Gateway
Hermes is not confined to a single terminal window. A single Hermes process drives 17 different platforms simultaneously. You can interact with the exact same agent state via Telegram, Discord, Slack, WhatsApp, Signal, SMS, and your CLI.
10. Real-Time Voice (/voice)
Available across the CLI, Telegram, and Discord. You trigger /voice and speak directly to the agent. When you are away from the keyboard or driving, you can push architectural updates or draft documentation verbally.
11. Cron + /webhook-subscriptions
This entirely eliminates the need for expensive Zapier subscriptions. You can use plain language to schedule background operations: “Every Friday 5pm, summarize this week’s GitHub commits, post to Slack #standups.” Combine this with webhooks, and platforms like GitHub, Vercel, or Stripe can push payloads directly to your agent’s DMs for zero token latency.
12. Custom Slash Commands (Skills)
This is the ultimate leverage point. Hermes ships with 100+ out-of-the-box skills triggered by autocomplete slash commands (like /architecture-diagram or /systematic-debugging). But the real power is building your own. You write a script that scrapes a specific data source, formats it, and executes a workflow. You map it to a custom command, and it runs flawlessly forever across any platform.
Look,
You have access to an agent with persistent filesystem memory, mid-flight steering, 17-platform ubiquity, dynamic API routing, and cron automation.
If it feels like it is underdelivering, it isn’t the software’s fault.
You never gave it the instructions it was built to execute.
In case we are meeting for the first time, come over here, it’ll be worth the roller coaster of articles that are gonna come up in the next few weeks.
And If you need help for analyzing UFC fights, please check out BoutPredict :)