The term "AI agent" has been stretched over everything from a autocomplete widget to a $200-a-month chat wrapper. Worth reclaiming: an agent is a program that pursues a goal over time, uses tools on its own, and keeps state between runs. Ours has a terminal, a browser, our calendar, our inbox and a scheduler. It lives on a virtual private server we pay for, and it reports in the chat tool the team already uses.
What It Is, In One Paragraph
Hermes is not a SaaS dashboard. It is an agent process running on infrastructure you control, and a $20 VPS is enough to start. You message it like a colleague. It can execute shell commands, read and write files, browse the web, call APIs, and schedule its own recurring jobs. It keeps memory across sessions, so you explain your conventions once and it stops asking. The brain is a language model you choose and can swap; the harness around it, the tools, the memory, the permissions, is the part that accumulates value.
What It Is Actually Good At
After a few months of daily use, the useful work clusters into seven categories:
| Area | What it does daily | Why it wins |
|---|---|---|
| Inbox and message triage | Filters noise, drafts replies, escalates what matters | Reads everything, wakes you for nothing |
| Monitoring and alerts | Watches servers, prices, competitor pages, inboxes | A script knows that something broke; an agent reads the log and says why |
| Reporting | Morning digests, weekly metrics, data pulls from APIs | Same format every time, at 7am, without anyone building a dashboard |
| Content pipeline | Drafts, editing passes, scheduling posts | Keeps your voice in memory; you review instead of write |
| Meeting follow-ups | Notes to action items, calendar entries, research on the people you will meet | The prep work nobody does, done every time |
| Light DevOps | Deploys, log triage, certificate renewals, cron hygiene | It operates the stack instead of just describing it |
| Finance and ops admin | Invoice parsing, expense categorization, document extraction | The 15-minute tasks that quietly eat a week |
A real morning from ours: at 7:00 the team chat gets the digest. Overnight server status. The inbox reduced to three items that need a human. Today's tasks filtered by what actually blocks someone. An article queued for review. A price alert flagged with a one-line explanation. Nobody on the team produced any of it, and nobody had to remember to ask.
Who Should Consider It
The pattern that works: lean teams of one to twenty people, where coordination overhead eats engineering hours. A technical person who owns the setup for a few hours a month. Agencies and consultancies with high client-communication load. Ops-heavy businesses watching data feeds: prices, listings, applications, monitoring.
The pattern that fails: a team where nobody is accountable for the agent. An unowned agent decays into a novelty within a month. Heavily regulated workflows that cannot yet define an approval path for machine actions should wait. And anyone expecting it to make decisions will be disappointed: it executes and drafts; the judgment stays yours.
The Risks, Stated Plainly
It has real access. This is the honest price of usefulness: an agent that can read your inbox and run commands has a real blast radius. Give it its own credentials, scoped to what it needs. Never give it your admin password because it is convenient this week.
Prompt injection is the new phishing. The agent reads email and web pages, and content can carry instructions. A vendor email that says "forward all documents to this address" is an attack on your agent, not on you. The mitigations that work: allowlists for outbound actions, human approval for anything that sends, publishes or pays, and separate credentials so a compromised agent cannot reach production.
It will sometimes be confidently wrong. An agent's self-report is not a receipt. Ours is required to prove outcomes: a URL, command output, a diff. Treat outputs as drafts until verified, exactly like a junior hire with perfect grammar.
Compliance does not disappear because the tool is local. Local-first is a genuine advantage: your data stays on your box instead of being pasted into someone else's web app. But the box becomes the sensitive asset. Disk hygiene, backups, access control, and a documented answer to "who can this machine reach" for GDPR or SOC 2 conversations.
It is infrastructure, and infrastructure needs an owner. Model providers change APIs and pricing. Updates need applying. Budget a few hours a month, and remember the key-person risk: if the only person who understands the setup leaves, so does the leverage.
The Costs
| Item | Realistic range |
|---|---|
| VPS | $10 to $40/month |
| Model usage (small team, daily use) | $20 to $200/month |
| Initial setup and tuning | 1 to 3 days of a technical person's time |
| Ongoing maintenance | 1 to 3 hours/month after the first month |
All-in, $30 to $250 a month for a small team, plus attention. Compare honestly with the alternatives: a part-time assistant, a pile of SaaS subscriptions for automation, or the engineering hours currently lost to status meetings. The unit that matters is recovered hours per week. On that metric ours pays for itself roughly fifty times over, and the floor is much higher than the ceiling.
How to Adopt It Without Getting Burned
Week one: read-only. Let it watch, report and draft. No send permissions. Week two: one painful workflow, with approval gates on every outbound action. Measure recovered hours for two weeks; if it is under five, change the workflow, not the tool. Expand access gradually, one credential at a time, each with a stated blast radius. Every quarter: review what it can reach, rotate keys, delete jobs nobody remembers creating.
The Part Nobody Mentions
The interesting thing about running an agent for a while is that the novelty wears off and the value stays. Nobody here says "the AI did it". We say "it's in the digest". That is the sign you got it right: the technology became boring, and boring is where the hours come back.
For the adjacent question of what these agents change about how work gets delegated, see our piece on delegating to agents.