Hermes Agent Not Responding

One test that halves the problem

Open the Hermes dashboard and ask the agent something trivial that needs no tools. What happens next splits the five causes into two groups, and saves you from debugging the wrong half.

• You get an answer. The instance is running and the model provider is accepting requests. Your problem is tool calling or something in the specific task, not the setup.
• You get an error. The provider is rejecting you: an invalid key, an empty balance, a rate limit or a retired model name.
• Nothing happens at all, or the page will not load. Either the instance is down or your route to it is. Those are different faults with different fixes, covered below.

Is the instance actually running?

Check the instance in your Flux dashboard rather than judging by the agent UI, because a container stuck in a restart loop looks intermittently alive. The usual causes are memory and a long-running task that never terminated.

Memory is the one worth checking first if you moved beyond light use. The 4 GB tier is comfortable for a personal agent doing scheduled and ad-hoc work, but concurrency is what consumes RAM: several tasks at once, a long multi-step research run holding a large working set, or a task that downloads and processes files. If your usage changed shape and the agent started dying, that is the 8 GB tier telling you it is needed rather than a defect.

A related pattern is an agent that hangs rather than crashes. An agent loop waiting on a tool that never returns, for instance a network call to a service that is down, will sit there indefinitely and look identical to a dead agent. Restarting the instance clears it, and the run history usually shows which step it stopped on.

The provider key, the balance, and the model name

Hermes is bring-your-own-key, so the most common silence originates at your AI provider rather than in the agent. Four failures produce the same symptom and all four are fixed on the provider side:

Re-paste the key rather than reading it and deciding it looks correct. A trailing space survives a visual check and fails every request.

Rate limits bite agents harder than chatbots

This deserves its own section because agents are unusually good at hitting limits. A chat assistant makes one request per human message. An agent makes a request per step of its reasoning loop, and a multi-step task can produce dozens of calls in a minute without you doing anything at all.

OpenRouter is a common choice for Hermes and its free tier makes the effect concrete. Free model variants, the ones whose identifier ends in :free, are capped at 20 requests per minute, and the daily allowance is 50 requests if you have purchased less than $10 of credits in the lifetime of the account, rising to 1000 per day once you have. An agent working through a research task can exhaust 50 daily requests in a single run, at which point it stops mid-task with no obvious explanation.

If your agent completes short tasks and consistently dies on long ones, this is almost certainly why. The fix is either a paid model variant, or accepting the ceiling and keeping tasks small. A one-off top-up is what moves the daily floor; it does not raise the per-minute cap, so a very chatty loop can still hit 20 per minute regardless of what you have spent.

The agent talks but never acts

This is the failure people rarely diagnose correctly, because nothing errors. The agent replies fluently, describes what it is going to do, and then simply does not do it. No tool runs, no file appears, nothing happens outside the conversation.

The cause is the model. An agent works by asking the model to emit a structured tool call, and not every model supports that. Point Hermes at a model without tool calling and you get a competent conversational partner that has been quietly stripped of its ability to act. It is not broken; it has no hands.

Check that the model you selected lists tool or function calling among its capabilities, and prefer a model the provider explicitly documents for agent use. This is also why the cheapest available model is often a false economy for an agent: capability at following a tool-calling protocol reliably matters far more than raw price per token when the whole point is that the agent executes.

A softer version of the same problem is a model that supports tool calling but does it unreliably, producing malformed calls that fail silently and get retried. If your agent half-works and the run history is full of repeated attempts at the same step, try a stronger model before assuming the task is at fault.

When the agent is fine and you cannot reach it

Hermes ships with a Tailscale mesh so the agent can reach your private network, and many people use that same mesh to reach the dashboard. That creates a failure mode where the agent is perfectly healthy, its scheduled tasks are still running, and you simply have no route to it.

Auth keys expire, and device keys expire too, so an instance untouched for months can drop off the tailnet with nothing having gone wrong. Check the Tailscale admin console: if the device is missing or unauthorised, generate a new auth key and reconnect. The giveaway for this cause is that scheduled work kept happening while the dashboard was unreachable.

The mirror image is worth knowing too. If the agent can reach the dashboard but cannot reach your internal services, the mesh is up for you and down for it, which usually means the device lost its authorisation rather than that the services moved.

The checklist

1. Ask the agent something trivial in the dashboard. An answer clears the instance and the provider in one step.
2. Check the instance in your Flux dashboard for a restart loop, and check RAM if your workload got heavier.
3. Check the provider dashboard for an invalid key, an empty balance, a rate limit or a retired model name.
4. If long tasks fail and short ones succeed, assume a rate limit and check your daily allowance before anything else.
5. If the agent talks but never acts, confirm the selected model supports tool calling.
6. If the dashboard is unreachable but scheduled tasks kept running, check Tailscale for an expired device or auth key.

Almost every case lands on steps one, three or five. An agent has more failure points than a chat assistant because it does more, and each one leaves a different fingerprint.