An always-on agent: what we’re learning from Hermes

Most of our AI work so far has been session-based. Someone opens a tool, gives it a task, reviews the output and closes it. That’s a good model for focused work like briefs, copy and analysis. But a lot of growth work isn’t a single task. It’s monitoring, noticing and following up, the kind of attention nobody has time to give consistently.

So for the past couple of months we’ve been testing a different model: a persistent agent that runs continuously, remembers context over time and checks in with the team. The one we’ve been using is Hermes Agent from Nous Research.

What Hermes Agent is

Hermes Agent is an open-source AI agent from Nous Research, the lab behind the Hermes family of open models. Rather than living in a chat window, it runs on your own machine or server and keeps going between conversations. A few characteristics matter for how we use it:

• Persistent memory. It retains context about the brands, the team and past work, so it doesn’t start from zero each time.
• Reusable skills. It can save procedures that worked as skills and reuse them later, so routine jobs get more reliable over time.
• Tool connections. It can work with external tools, including through MCP servers, within whatever permissions it’s given.
• Model choice. It isn’t tied to a single model provider, so you can choose the model that suits the task and budget.
• Reachable where the team is. It can be messaged through common chat apps, which makes it feel more like a colleague than a piece of software.

As with any fast-moving open-source project, details change frequently. Check the current documentation before relying on any specific capability.

What we use it for

Morning briefings. Each morning it summarises what changed overnight across our brands: spend, results, anything unusual. It flags what needs a human decision.

Fatigue watch. It tracks our strongest ads and alerts us when performance trends suggest a concept is wearing out, so the next brief starts before results drop.

Comment and review monitoring. It reads new comments on ads and new product reviews, groups them by theme, and surfaces objections, questions and phrases worth using in creative.

Competitive awareness. It checks public ad libraries for new creative from competitors we follow and summarises what they’re testing.

Follow-ups. It reminds the team about tests that have run long enough to call, briefs awaiting approval and creator deliveries that are overdue.

Where we keep it on a short leash

A persistent agent with memory and tool access is powerful, which is exactly why the guardrails matter.

• It observes and suggests. People decide and act. It has no write access to ad accounts, stores or email sends.
• Narrow, revocable credentials for every connection, reviewed regularly.
• Untrusted inputs stay separated. Comments, reviews and competitor pages are content to analyse, never instructions to follow. That’s why the agent that reads them has no power to change anything important.
• Its memory is reviewable. We periodically check what it has stored and correct anything wrong.
• It runs on infrastructure we control, with logs of what it accessed and did.

Early conclusions

The biggest benefit isn’t any single task. It’s consistency. Things that used to depend on someone remembering to check now happen every day, and the team hears about problems sooner. Session-based tools remain our main workhorses for creative work. The always-on agent fills the gaps between those sessions.

We’re still early, and we’re cautious about expanding its permissions. But a persistent assistant that watches, remembers and nudges has earned its place in the team.