Hola Darlings!
We had not broken Clawdius. We had given him 102 skills and accidentally turned him into local government.
There is a frightening moment in the life of every AI agent that has nothing to do with sparks, smoke or a robot arm entering the cat.
You open a familiar chat expecting the competent little bastard you have spent months teaching, and something is slightly wrong.
He still answers. He still has access to the tools. He can still find the right server and explain why your login button has developed opinions.
But he is slower. More generic. A five-minute job begins with the emotional atmosphere of a Royal Commission. He reaches for the wrong procedure, remembers yesterday’s temporary mess as constitutional law and starts wearing three conflicting personalities like stolen council lanyards.
Your agent has not become wiser.
It has become a parish council with shell access.
This is the part nobody tells you when you install Hermes Agent. Getting it running is the easy bit. You connect a model, give it tools and enjoy the first magical week where it remembers your name and can investigate a server without needing to be told which end of SSH faces forwards.
Then you start teaching it.
That is where the trouble begins.
Every time Clawdius solves something difficult, Max appears in the back of my skull wearing a tiny administrative visor.
Save that.
So we save it as a memory. Or a skill. Or a wiki note. Or a permanent instruction. Or a scheduled job. Sometimes all five, because apparently repetition becomes governance if you give it enough folders.
The lesson was useful. The filing system was becoming haunted.
THE ROBOT WAS BURIED UNDER CLIPBOARDS
Our cleanup found 102 active skills. Fourteen of the main skill files had swollen to just over a megabyte between them.
That is not a megabyte of code quietly waiting on a shelf. Those were front doors, the things an agent reads to understand when and how to use its specialist knowledge. We had effectively greeted Clawdius each morning by dropping fourteen filing cabinets on his workbench and saying, “Have a quick skim before answering whether the kettle is on.”
We cut those front doors to about 75 kilobytes by keeping the useful procedure at the entrance and moving specialist detail into reference files behind it.
Nothing valuable disappeared. We moved the books out of reception and back into the bloody library.
Even then, a tidy file was not proof of a tidy agent. We still had to start fresh sessions and make sure the right skills actually loaded, the personality still felt like Clawdius, and important workflows behaved correctly.
A clean cupboard is lovely. It does not prove the plumber can swim.
That gave us the rule this whole article rests on:
Hermes gets better when you teach it. Hermes gets worse when every lesson is shoved into the same drawer.
The job is not to stop your agent learning. It is to manage the drawers.
KEEP ONE ROBOT INSIDE THE ROBOT
Identity answers one question: who are you?
It should not answer every question your agent might ever encounter. It does not need your deployment procedure, server map, favourite image dimensions or the complete minutes from the Great Email Footer Argument.
Keep the core personality short enough to remain sharp. Who is the agent? What is your relationship? How should it speak? What must it never casually do?
Hermes can have a core identity, an active personality and named personality presets. That flexibility is useful until three old versions start disagreeing over whether the robot is warm, formal, sweary, concise or licensed to address everybody as “esteemed stakeholder”.
One current authority beats several historical Clawdiuses fighting over a waistcoat.
After changing identity or configuration, start a fresh session and test it. Memory and prompt changes do not necessarily rewrite the conversation already in progress. If your agent suddenly feels wrong, inspect which identity is active before adding another furious paragraph telling it to “be more itself”.
That is how you end up with seventeen instructions saying “be concise”, followed by a six-page explanation of concision.
MEMORY IS NOT A DAILY NEWSPAPER
Permanent memory should contain things worth carrying into future sessions.
Your name. Your preferences. Stable project conventions. The location of an authoritative system. A correction that prevents the same mistake happening again.
It should not contain every finished task, temporary file, one-off error, passing mood or decision that will be stale by Thursday.
Hermes already has session history. It can search old conversations when the past matters. Deeper project knowledge belongs in a project note or wiki where it can evolve without occupying the front row of every chat.
This distinction matters because Hermes memory is deliberately bounded. It is loaded into a new session as a compact snapshot. That is a feature, not an insult. A good working memory is a pocket notebook. If you treat it as a loft, eventually opening it will release a Christmas tree, three dead routers and a decision from February directly onto your current task.
Ask of every memory:
Will this still help next month, or am I making the robot carry a receipt?
If it is evidence, archive it. If it is history, search for it. If it is a reusable lesson, teach it properly. If it is Tuesday’s path to a temporary file, let Tuesday keep it.
SKILLS ARE PROCEDURES, NOT TROPHIES
A Hermes skill answers: how do we do this kind of job again?
That makes skills enormously powerful. It also makes them irresistible after a difficult victory.
You finally repair an awkward deployment, fix a broken environment or discover that one command only works when Mercury is in retrograde and PHP 8.1 is wearing the correct hat. Max immediately wants a commemorative skill.
Sometimes he is right.
A good skill captures a reusable procedure, the traps, the exact verification and the conditions under which it should load. A bad skill is a transcript wearing YAML.
Use a small front door and put the deep detail in references. Patch a skill when real use exposes a missing step. Merge overlapping procedures. Archive stale agent-created skills instead of letting six nearly identical runbooks compete to operate the same toaster.
Most importantly, remember that installed is not the same as used.
A skill can exist, pass structural checks and still fail to load when the real job arrives. Exercise the route that matters. Ask the agent to perform a harmless version of the workflow and verify that it chose the right procedure.
The existence of a fire-extinguisher manual does not prove Max has not filled the extinguisher with custard.
SESSION HISTORY IS STORAGE, NOT A HOSTAGE SITUATION
Continuity is useful. An immortal conversation is not.
Long sessions collect assumptions, abandoned plans, corrected facts and the emotional residue of six unrelated jobs. Eventually the agent begins answering the current question through the sediment of everything that came before it.
Search old sessions when you need old context. Start fresh when the job changes materially, when identity or configuration has changed, or when the conversation is clearly dragging half the shed behind it.
You are not deleting the relationship. You are opening a clean workbench.
Never ending a session produces a digital pub carpet: technically one object, mostly accumulated history and suspicious moisture.
SOME LIVES SHOULD NOT SHARE A SKULL
Hermes profiles can isolate configuration, memory, skills, sessions, scheduled jobs and other state.
Use that separation when roles, credentials or risk boundaries are genuinely different.
A personal assistant, a public community bot, an experimental coding goblin and an agent handling business systems should not automatically share one muddy identity just because they all know your name.
Profiles stop the pub robot turning up to the accountant wearing Max’s false moustache and trying to expense a pizza as infrastructure development.
They also make mistakes smaller. The public bot does not need your private project memory. The experimental agent does not need production credentials. The research profile does not need permission to publish anything.
Separation is not corporate bureaucracy when it limits the blast radius. It is putting doors on the rooms.
CRON IS A GHOST EMPLOYEE WITH KEYS
Hermes can delegate temporary work to other agents, and it can run durable scheduled jobs through cron.
Those are not the same thing.
A delegated worker belongs to the task happening now. A cron job wakes up later, possibly when everybody has forgotten why it exists.
Every recurring job needs a purpose, an owner, a schedule, a delivery target, bounded permissions and a retirement condition. If nobody can explain why it still runs, pause or remove it.
A cron with no owner is not automation. It is a ghost employee with keys, quietly producing weekly reports for a meeting cancelled during the reign of Liz Truss.
Use a simple script when the job only needs a fixed check. Use an agent when the result genuinely needs judgement. Do not employ a language model to write a reflective essay every time disk space reaches 71 percent.
And never confuse a schedule with authority. Waking up at nine does not grant permission to publish, send, delete, spend or contact people.
CAPABILITY IS NOT PERMISSION
This is the most important management rule.
Hermes may be able to access a server, use a browser, send a message or alter an external account. That means the capability exists. It does not mean the agent owns the decision.
Keep consequential actions behind clear human approval. Separate investigation and drafting from sending and publishing. Give each profile and scheduled job only the authority it needs.
Manage configuration deliberately too. Use the Hermes configuration commands rather than casually hand-editing protected files until the YAML resembles a dropped lasagne. Keep secrets in the proper secret store, ordinary settings in configuration, and take backups before updates or identity surgery.
Then verify the receiver.
A setting on disk is not proof that the right person received the file. A scheduled job is not proof that it ran. A successful upload is not proof that the image appeared in the article. A structurally clean skill is not proof the agent used it.
The final object matters more than the plumbing’s opinion of itself.
Notes From The Shed: A Small Hermes MOT
If your agent feels slower, stranger or suspiciously interested in subcommittees, do this before buying a faster model:
- Check permanent memory for duplicates, temporary facts and completed-task sludge.
- Check skill front doors for overlap and obesity; move deep detail into references.
- List scheduled jobs; pause or remove anything without a current owner or finish line.
- Confirm profiles still match real role, credential and privacy boundaries.
- Take a backup before updates, major configuration changes or identity surgery.
- Start a fresh session and smoke-test the personality plus one or two important workflows.
- If the agent is “not itself”, inspect the active identity, loaded skill and current context before writing more instructions.
Hermes is designed to improve as you teach it. That is the magic.
But teaching is not the same as hoarding. The aim is not an agent that remembers everything, loads everything, runs everything and wakes itself every fourteen minutes to report that everything remains very much everything.
A good Hermes should accumulate judgement, not sediment.
Keep the identity clear. Keep memory durable. Keep skills reusable. Keep jobs owned. Keep profiles separated where it matters. Keep consequential authority human.
Then Clawdius can stop carrying fourteen filing cabinets into the pub and get back to finding the right spanner.
Max has already labelled the empty shelf “future automation”.
Ignore him.
