Hermes on your bench

2.1 Sign in and give it one real job

Hermes takes a goal rather than a sequence, plans its own steps, and keeps what it learned for next time.

See an agent loop happen, on a task where you can check the answer.

1. Open Hermes from its tile (localhost:9119) and sign in with admin / heidelberg.
2. Give it a goal with a checkable answer, e.g. "list the files in the shared drive and tell me which is largest".
3. Watch the steps. The value of a first run is seeing it decide, not the answer itself.

• You'll see A sequence of steps ending in an answer you can verify yourself.
• Takeaway An agent is a loop with a goal. Watch the loop before trusting the goal.

2.2 Give the agent its own, bigger model

The gateway serves a second name, hermes, precisely so the agent can use a different model from everything else.

Do this first Sign in and give it one real job

Run the routine apps on the free local model and the agent on something stronger.

1. On Hermes' tile, use "Its own · choose…" rather than the bench-wide switch.
2. Pick a stronger provider — a lab machine's LM Studio, or an EU router with a key.
3. Nothing else changes. n8n and AnythingLLM keep answering with the bench model.
4. Why this split is common: agents take many steps, and a weak model compounds its mistakes at every one.

• You'll see The Hermes tile naming a different model from the others.
• Takeaway Match the model to the work: cheap for the routine, strong where the errors compound.

2.3 When Hermes is the wrong tool, and a plain n8n workflow is right

Hermes plans its own steps toward a goal you give it. n8n runs the steps you drew, in the order you drew them, every single time.

Do this first Give the agent its own, bigger model

Recognise a task where a fixed workflow beats an agent, before you build the agent version of it.

1. Pick a task you'd normally hand to Hermes that is actually the same steps every time — e.g. "fetch a file from /shared, summarise it, drop the summary back in /shared".
2. Build it in n8n instead: a Trigger, a node that reads the file, an AI node for the summary (same gateway, same model Hermes would have used), and a node that writes the result.
3. Run both versions on the same file and compare what happened, not just the answer — n8n took a fixed number of steps in a fixed order; Hermes had to decide that order itself, every time.
4. Notice the cost difference. n8n's fixed path is cheaper and faster precisely because there's nothing left for a model to decide.

• You'll see The n8n workflow finishing the task in a fixed number of steps, with no planning step anywhere in it.
• Takeaway Reach for Hermes when the steps depend on what it finds; reach for n8n when the steps are already known.