Spinning up Hermes Agent on an exe.dev VM took a few minutes. Wire it to Telegram and the bot lives in my pocket. A week in, 213 sessions and 7,263 messages later, this is what has stuck.
What I actually use it for
The killer feature is not the chat. It is that the chat compounds.
When I ask the agent to do something twice, I tell it to remember how. It writes a skill: a markdown file with a name, a description, and the steps it just figured out. Next time I ask, it loads its own notes. Sometimes the agent offers the skill on its own after a hard task. Sometimes I just describe a workflow in chat and tell it to save that as a skill before I ever run it.
One that has earned its keep: flying. I am a student pilot at a flight school whose mobile app has no public API. I asked Hermes to reverse-engineer it. Now a skill called aviatize-operator books, cancels, and reschedules my lessons. Another, aviatize-watcher, polls every fifteen minutes during hours I would actually fly and pings me when an instructor opens a slot.
Not impressive on its own. The point is that it did not exist when I started.
What is still rough
A few edges.
Memory. Hermes auto-trims and updates the memory files it injects on every turn, which works well enough for now. I want to try the pluggable memory providers (Honcho, mem0, supermemory, and the others) to see what changes when memory lives in a real backend instead of a capped prompt slot.
Backup. Everything important (skills, memory, cron jobs, the chat history) lives on one VM. If the machine goes away, so does a week of accumulated context. I have not picked a backup strategy yet.
Letting the agent touch GitHub safely. I run two accounts: my personal one and a bot personal account. The bot opens PRs and pushes branches; my personal account reviews and merges. They cooperate through narrow personal-access tokens scoped to specific repos. It works, but the wiring is more fiddly than I would like: token rotation, fork sync, branch protection rules.
Scoped access to Google. I want the agent to read a few calendars and a few Drive folders, not all of them. Google's consent model is closer to all-or-nothing than I would like, and the per-document sharing route does not compose well when the agent is the one doing the sharing.
Sharing things I do not want to paste. PDFs, screenshots of bank statements, the half-written letter I am asking it to edit. Anything I would rather not drop into a chat window. The agent can read files from disk just fine. Getting them onto the disk from my phone, without the contents passing through somebody's chat history, is the friction.
The bigger question
How much to trust the agent in the seat I am not in.
Some work is read-only and easy to delegate. Some work touches money or calendars or other people. Most work lives in the gap between those two, and I do not have a clean answer for what belongs where. Sometimes a confirmation prompt. Sometimes a dry-run mode. Sometimes a separate account. Sometimes paying closer attention. I am still figuring out the shape.
Hermes Agent: NousResearch/hermes-agent.