agentos.guide › blog
A hermes agent fallback model keeps every profile in your Agent OS answering when its main provider stalls: you list one or more backup provider:model pairs per profile, and Hermes swaps to them automatically on rate limits, outages and failed logins.
This guide sets that up profile by profile, shows how to prove each chain works, and finishes with three prompts that get the agent to audit and maintain its own backups.
The keys and commands come from the official Hermes Fallback Providers page.
If you'd like help planning which profiles need what, book a free strategy session with my team.
An Agent OS runs several Hermes profiles side by side, each with its own config.yaml.
Fallback is configured per config file, so a research profile and an operations profile can carry different backups.
That's useful. A profile drafting long reports may want a backup with a big context window, while a quick-reply support profile just needs something fast and reliable.
Members of the AI Profit Boardroom get the installable Agent OS, so the profiles already exist before you add backups to them.
Use the interactive manager with the profile flag:
hermes -p research fallback add
hermes -p research fallback list
The picker is the same one hermes model uses. Pick a provider different from the profile's main one, sign in, choose the model.
Behind the scenes it writes a top-level fallback_providers list:
fallback_providers:
- provider: openrouter
model: anthropic/claude-sonnet-4
Both lines are required for every entry. If you inherit an old profile with the singular fallback_model key, the manager migrates it the next time it writes.
| Failure | When Hermes switches |
|---|---|
| Rate limit (429) | Once retries are used up |
| Server error (500, 502, 503) | Once retries are used up |
| Login rejected (401, 403) | Straight away |
| Model not found (404) | Straight away |
| Repeated empty or broken replies | After the repeats |
The swap happens mid-conversation. History and tool calls carry over, and the backup gets its own reasoning effort setting.
It lasts a single turn: the next message goes back to the primary, unless the primary has reported a reset time that hasn't arrived yet.
🔥 Want profiles that come with sensible backups?
Inside the AI Profit Boardroom you get the installable Agent OS, step-by-step trainings, four coaching calls a week and 3,400+ members sharing their profile configs.
Failover is tied to the turn, not the profile. Every new message to a profile tries its primary first, and within a turn the backup is used once at most.
If the backup fails as well, the normal retries run and then you get an error, so two broken providers can't trap a profile in a loop.
When the primary says when it will reopen, for instance at the end of a subscription window, Hermes keeps that profile on the backup until the time passes rather than retrying a request it knows will fail.
To wait out short windows instead of switching, set fallback.min_switch_reset_seconds in that profile's config. It defaults to 0, which means always switch.
Bear in mind each switch empties the prompt cache for that conversation, so a profile that keeps flipping between providers re-reads its whole history every time.
An Agent OS leans on work that happens out of sight, so check three places.
auxiliary. On provider: auto they try the main model, then any task fallback_chain, then your profile's fallback_providers.cron.model and cron.model_provider if you want failover.delegation.fallback_providers when it's set; otherwise only unpinned children inherit the parent's chain.Pinned helper tasks behave differently: they only move on for capacity problems, such as a payment error or an exhausted daily quota, not a short rate limit.
The readiness check, listed in the v0.23.2 release notes, tests a chain without opening a session:
hermes -p research fallback check --live
hermes -p support fallback check --live --json
Each entry gets one small request on its own provider. You'll see statuses such as ready, auth, model_mismatch or live_timeout.
It exits 1 when anything isn't ready, so a scheduled job can run it across all your profiles and message you when a backup breaks.
Credentials are never printed, which makes the output safe to send to a chat.
Run the plain version first if you only want to confirm the config and logins resolve; it sends no model request at all. Add --live when you want proof that each backup actually answers.
Plenty of Agent OS owners run a model on their own machine for everyday profiles and keep a hosted model as the safety net.
The providers documentation shows a local Ollama primary set through the custom provider with a localhost base_url; you then add a hosted entry to fallback_providers.
Give the local model at least 64,000 tokens of context, the minimum Hermes accepts for agent work with tools.
If the local server is switched off or overloaded, the hosted backup takes the turn, and the next message tries your own machine again.
For choosing that local model, see our guide to the best local model for Hermes Agent.
One limit: core Hermes doesn't ask before moving to a paid backup. A confirm option exists only as an open proposal (PR #149020).
Paste these into the profile you want to work on. Each one asks the agent to show you changes before making them.
Read ~/.hermes/config.yaml for this profile. List my primary model and
provider, every entry in fallback_providers, any auxiliary task with its own
fallback_chain, and any cron job pinned to its own provider or model.
Flag any backup on the same provider as the primary and any entry missing
a provider or model. Then run: hermes fallback check --live
Summarise the status of each entry in a short table. Change nothing.
My primary provider is in config.yaml. Suggest one backup model from a
DIFFERENT provider that I already have credentials for, that supports tool
calling and has at least 64,000 tokens of context. Show me the exact
fallback_providers entry you would add and wait for my yes. After I approve,
add it, then run hermes fallback check --live and report the result.
Create a weekly cron job for this profile that runs
hermes fallback check --live --json and sends me a short message only when
an entry is not ready, naming the entry and its status. Do not pin the job
to a provider or model. Show me the job before you create it.
model-fallback-heuristic is a community plugin by tomkubik, version 0.2.0, in the Hermes Marketplace. It isn't from Nous Research.
It ranks models from public catalogues and rewrites fallback_providers at session start and after errors, keeping entries you wrote yourself. It leaves the primary alone unless you opt in.
It suits a profile that runs mostly on free models. For a steady paid profile, the core chain plus the weekly check is simpler.
📞 Want your profiles' backups planned with you?
Book a free strategy session and we’ll go through each profile, its main model and the right spare.
config.yaml or the hermes fallback command only.Give every customer-facing profile one backup on a different provider, prove it with the live check, and let a weekly job watch it.
Leave scheduled jobs unpinned so they inherit the chain, and add helper-task chains only where a specific model matters.
If you'd like the whole Agent OS installed with this already thought through, the AI Profit Boardroom includes it.
No. Each profile has its own config.yaml, so each can carry its own fallback_providers list. Put hermes -p and the profile name in front of each fallback command.
Run hermes -p with each profile name, then fallback check --live, or schedule a job that does it and only messages you on failures.
Yes, if it is unpinned. A job given its own provider or model will fail instead of switching.
It can edit config.yaml if you allow it. The prompts above ask it to show the change and wait for your approval first.
Not in core Hermes yet. A confirm option has been proposed in an open pull request.
Only if a profile rotates between many free models. It is a community plugin; the built-in chain covers most Agent OS setups.
Run the audit prompt on one profile today, fix what it flags, and if you want the full Agent OS ready to install, the AI Profit Boardroom includes it, and a free strategy session with my team can map out each profile’s hermes agent fallback model.
About Julian
I'm Julian Goldie, an AI entrepreneur, SEO expert and the founder of the AI Profit Boardroom, which has 3,400+ members.
I help business owners scale with AI agents, automation and SEO.
My YouTube channel has 400,000+ subscribers, and I run Goldie Agency, a seven-figure SEO agency.
→ Hermes Agent Cron Job: Schedule, Stop And Recover Jobs In Your Agent OS With 5 Copy-Paste Prompts
→ Best Local Model For Hermes Agent (And The Exact Setup)
→ Hermes Agent Plugins Marketplace: The Quick Start, And Where Each Plugin Fits In Your Agent OS
→ Hermes Agent Subagents: The Quick Start, And How They Fit Your Agent OS
The installable system behind every guide on this site — dashboard, agents, memory and pipelines — updated as new tools land, with 3,400+ members and four coaching calls a week.
Join the AI Profit Boardroom →📺 Video notes + links to the tools 👉
🎥 Learn how I make these videos 👉