# Hermes Agent users choose their models, and Grok Bot users ask to do the same

*In seven weeks of posts, 128 people described choosing, switching or mixing the models under Hermes Agent, which is 1st of 5 on model choice. In the same weeks 57 users of Grok Bot, which has no model picker in its settings, asked to pick, pin or see it, and some named Hermes as the example.*

What users credit, Hermes Agent. By Feedback Bench, 17 October 2026. Feedback Bench 1.1. Posts 2026-08-31 to 2026-10-18; bench built 2026-10-19. Web: https://feedbackbench.com/#/articles/hermes-agent-users-choose-their-models-and-grok-bot-users-ask-to-do-the-same

> one of the benefits of running a hermes agent is being able to switch models on the fly. gpt and claude are down for us right now. i hit /model and switch to grok 4.6 on the fly (not down for us right now) and am able to continue our daily workflows.
> the model switching is seamless because the hermes agent is connected to our company brain with extensive session notes from what we’ve already been working on today.

A user on X, 3 September ([post](https://twitter.com/375707743/status/2095533246519877651))

That day OpenAI's status page logged elevated errors across ChatGPT and Codex, and Anthropic's logged elevated errors for several Claude models; users on X said Grok was down too. This poster, writing for a team, put a different model in front of the same day's notes with one command. Another came back once the services had recovered. "[my work never stopped, though. i switched my hermes agent to glm and qwen3.8-max and kept going.](https://twitter.com/895324925380235264/status/2095561660731634156)"

Five of the people we counted describe a moment like this, when a provider went down, a usage limit ran out or an account was banned. Four of them switched by hand, and one had a fallback set up in advance.

## The strength

Hermes users choose the model under their agent and change it when they need to. Across the seven weeks, 128 people described doing this first-hand; to be counted, a person had to say they switch, mix, route or fall back between models inside Hermes, or that Hermes does not tie them to one provider. On [Which AI model runs, how good it is, and choosing it](https://feedbackbench.com/criteria/control.model.md) Hermes Agent is 1st of 5, and its interval sits clear of OpenClaw in second place. Using the bench's own labels with one rule for every agent, 5.9% of the people judging Hermes praised its model choice, against 2.4% for OpenClaw and under one percent for everyone else.

**Exhibit 1: Hermes Agent users praise model choice at more than twice the rate of the next agent.** Share of people judging each agent whose posts praise its model choice, after exclusions

|  | Share of judging people |
|---|---|
| Hermes Agent | 5.9% |
| OpenClaw | 2.4% |
| ChatGPT Dots | 0.9% |
| Gemini Spark | 0.8% |
| Grok Bot | 0.6% |
| Muse | 0.4% |
| Instinct | 0.0% |
| Poke | 0.0% |

*Same labels and rule for every agent. Collection differs: Muse's base includes app-store reviews (0.9% without them), Gemini Spark is X and YouTube only, Poke is one maker handle. Only Hermes Agent had an extra referral screen, which lowers its rate slightly. n = 49,400 judging people, 6,926 on Hermes Agent. Source: Feedback Bench labels. Posts 31 Aug to 18 Oct 2026.*

## The users who asked for it elsewhere

The clearest measure of what Hermes users have came from people who did not have it. Grok Bot, xAI's personal agent, has no model picker in its settings, and its users asked for that choice all through the window. We read every Grok Bot post our keyword search found on choosing or seeing the model. Across the seven weeks, 57 Grok Bot users asked first-hand to pick, pin or see the model, or complained that they could not. Thirty-seven wanted to pick or pin a model; the other 20 wanted to know which one had run.

Many asked early and liked the product. Early in September one user, who runs a small swarm of Grok Bot agents and loves the product, listed what was missing. "[grok is still not the best underlying model for everything. hermes is cool because you can switch the model depending on the agent's task.](https://twitter.com/1805382710757834752/status/2096289186466939110)" Late in the month another wrote, "[one thing i’m missing with @bot, which i generally really like, is being able to switch models.](https://twitter.com/748776072477487104/status/2104358301542621472)"

On 7 October Elon Musk [posted on X](https://x.com/elonmusk/status/2107724314451878104) that, going forward, Grok Bot would use the best backend model for each task, naming Claude Opus 5.5, Midjourney and Suno among others. Twenty-seven of the 57 first asked from 7 October, the day of the post, and just over half of those asked to see which model had run rather than to choose one. One welcomed the move in a reply to Musk, then asked, "[could we get an option so grok bot doesn't pick the model, and instead we can pin one, for example claude opus 5.5, and always use it for every task?](https://twitter.com/1690112069641646080/status/2107801751785021805)" A new user who had just cleaned up their Gmail with a team of bots wrote, "[one wish: show me which model is doing the work.](https://twitter.com/2082536149458046976/status/2109288148488618264)" Days after the post, one user, not among the 57, put it to Musk directly. "[are you able to make grok bot offer model selections like hermes agent does???](https://twitter.com/1086080362038476802/status/2109051364202672219)"

The bench's own request extraction points the same way. On this criterion, manual choice of the model is the most requested change, with 98 user-weeks; Grok Bot accounts for 54 of them and OpenClaw, next, for 22. Per thousand people judging each agent, Grok Bot users ask for choice about twice as often as Hermes users do.

Not every Grok Bot user wants a picker. "[i dont care what model is being used as long as it does what i need.](https://twitter.com/2012613091339460609/status/2107759343441695074)" one wrote on the day of the post. In the five days after the post, eighty-three Grok Bot users welcomed the routing, against twenty-one who asked to see, pick or pin the model and seven who asked to opt out of other makers' models. Some want the opposite of choice, a router that sends simple work to cheaper models.

**Exhibit 2: Per person, Grok Bot users ask more often to choose the model, and Hermes Agent users to have it routed.** User-weeks asking for each change, per 1,000 people judging the agent

|  | User-weeks per 1,000 judging people |
|---|---|
| Let me choose the model: Hermes Agent | 1.7 (12 user-weeks) |
| Let me choose the model: Grok Bot | 3.8 (54 user-weeks) |
| Route tasks to models for me: Hermes Agent | 2.2 (15 user-weeks) |
| Route tasks to models for me: Grok Bot | 0.4 (5 user-weeks) |

*From the bench's request extraction on the model-choice criterion: manual choice, Hermes Agent 12 and Grok Bot 54 user-weeks; automatic task-based routing, Hermes Agent 15 and Grok Bot 5. Divided by each agent's judging people after exclusions. n = 6,926 Hermes Agent and 14,182 Grok Bot judging people. Source: Feedback Bench requests. Posts 31 Aug to 18 Oct 2026.*

## What Hermes users ask for next

Hermes users, who already have the picker, ask for something smaller and different. Sixteen in our reading ask for routing, a cheap model for routine work and a stronger one when a task gets hard, or have built it themselves. A heavy user running several agents and scheduled jobs put the question to r/hermesagent. "[and has anyone got a good automatic model routing setup working? something like cheap model for normal tasks → stronger model when things get difficult.](https://www.reddit.com/r/hermesagent/comments/1wr5jka/what_models_are_you_guys_actually_using_with/)" Another described the same split for subagents: "[a cheap/fast model for exploration or implementation, a stronger model for review, dispatched from a main thread.](https://www.reddit.com/r/hermesagent/comments/1wp7foj/how_do_u_run_subagents_with_a_specific/)" Per person, Hermes users ask for routing about six times as often as Grok Bot users.

## What the source shows

Hermes's code explains why the switch in the opening works. A typed `/model` changes the model inside the running session. In `hermes_cli/cli_model_switch_mixin.py` the conversation history is kept, and the next user message carries a note telling the new model that it has taken over and should "Adjust your self-identification accordingly." If the new model's context window is smaller, a preflight check warns about compression. By default the switch lasts for that session only (`persist_switch_by_default` in `hermes_cli/model_switch.py`). Memory sits in the profile's own home folder and is injected each turn by `MemoryManager` whatever the provider.

Automatic fallback is optional and works turn by turn: the docs say it "automatically switches to a different provider:model when your main model fails". Since mid-September, a turn that fails after retries ends by suggesting `hermes fallback add`.

Routing has a longer history, and Nous has taken positions on it. An opt-in smart routing feature that sent short, simple turns to a cheap model was removed in a pull request merged on 20 April; the request states no reason. A fix filed days earlier showed that it could silently override a model the user had chosen with `/model`. The `model` parameter on `delegate_task` was removed in March, which a later docs change describes as intentional, "to enforce consistent delegation behavior". In July the project's lead maintainer closed one of many pull requests to add a per-task model with the words "We do not want this". Since August the docs describe a cheaper path, a strong main model with an inexpensive `delegation.model` for every subagent, and that setting has been in the model picker since 17 August. A user who tested subagents described what that leaves out.

> overall solid: subagents get isolated context + tools and just work. one gap: no per-task model — delegation.model/provider in config.yaml pins all subagents globally, so mixed-model batches still require separate hermes --provider/-m processes

A user on X, 10 October ([post](https://twitter.com/2077810080645263361/status/2108945297346842683))

The question is still open. The project's triage notes say maintainers need to choose one design before any per-task proposal can land, and they point to an operator-defined list of named models. Meanwhile the plugin catalog carries community routers.

## Who gains most

People who run several agents gain most. Thirty-two of the 128 also post about OpenClaw, and at least four say they came from it, among them a user who wrote, "[giving hermes @nousresearch another shot after a pretty poor experience using openclaw at the start of the year. i love that it's model agnotistic and i can use my @claudeai sub.](https://twitter.com/1457145836878458881/status/2106516968161554477)" Hermes ships an OpenClaw migration in its optional skills. Ten of the 128 measure Hermes against Grok Bot's model, which users cannot choose.

## Two things to build

**Make the fallback suggestion one step.** Four of the five people who lived through a failure switched by hand. A failed turn already suggests `hermes fallback add`. Hermes also knows which providers a user has signed in to (`get_authenticated_provider_slugs`), so the suggestion could name one of them and add it in a single step.

**Use the routing requests as evidence for the open design.** The users asking for routing describe rules they set themselves, not a router that guesses. That matches the operator-defined list of named models the triage notes point to, and not a free-form model id the agent picks. The removed router is a reason to keep the user's pin as the default, since it could override a `/model` choice without telling anyone.

## Limits

Our routing and picker counts come from reading posts found by keyword, so they are floors. We left out posts promoting one routing product. Grok Bot has about twice as many judging people as Hermes, so the comparisons above are per person. On model choice itself, some users call the Hermes picker confusing, and a few report that a model which does well elsewhere does worse inside Hermes; 167 people complained on this criterion against 408 who praised it.

## How we read this

We read public posts about eight personal agents over the seven-week window, using Feedback Bench 1.1. AI agents read and classified the posts under written rules, and the bench labels come from large language models. AI agents read every Hermes post the bench labelled as praise for model choice, and every Grok Bot post that mentioned choosing or seeing a model, and counted the people under rules written beforehand, leaving out promoters and anyone on our exclusion list. The cross-agent rate uses the bench labels with one rule for all eight agents, and only Hermes got an extra referral screen, so its rate is, if anything, slightly low. We read the Hermes source, docs, pull requests and issues on 17 October.

## Sources

- Feedback Bench, [Which AI model runs, how good it is, and choosing it](https://feedbackbench.com/criteria/control.model.md) and [Hermes Agent](https://feedbackbench.com/agents/hermes-agent.md), built 2026-10-19.
- Elon Musk on X, [post of 7 October 2026](https://x.com/elonmusk/status/2107724314451878104), read 2026-10-17.
- Nous Research, [Hermes Agent README](https://github.com/NousResearch/hermes-agent), read 2026-10-17.
- Nous Research, [Fallback providers documentation](https://hermes-agent.nousresearch.com/docs/user-guide/features/fallback-providers), read 2026-10-17.
- Hermes source, [hermes_cli/cli_model_switch_mixin.py](https://github.com/NousResearch/hermes-agent/blob/main/hermes_cli/cli_model_switch_mixin.py), [hermes_cli/model_switch.py](https://github.com/NousResearch/hermes-agent/blob/main/hermes_cli/model_switch.py), [agent/memory_manager.py](https://github.com/NousResearch/hermes-agent/blob/main/agent/memory_manager.py), [agent/turn_failure_copy.py](https://github.com/NousResearch/hermes-agent/blob/main/agent/turn_failure_copy.py), read 2026-10-17.
- Hermes pull request [#12732](https://github.com/NousResearch/hermes-agent/pull/12732), removal of smart model routing, merged 2026-04-20; [#8459](https://github.com/NousResearch/hermes-agent/pull/8459), fix for routing overriding `/model`, closed 2026-04-21.
- Hermes pull request [#49489](https://github.com/NousResearch/hermes-agent/pull/49489), delegate_task model is config-level, merged 2026-06-20; [#20000](https://github.com/NousResearch/hermes-agent/pull/20000), per-task model routing for delegate_task, closed 2026-07-05; read 2026-10-17.
- Hermes pull requests [#83062](https://github.com/NousResearch/hermes-agent/pull/83062), frontier planner and inexpensive worker docs, merged 2026-08-10, and [#80463](https://github.com/NousResearch/hermes-agent/pull/80463), subagent model in the picker, merged 2026-08-17.
- Hermes pull request [#87179](https://github.com/NousResearch/hermes-agent/pull/87179), operator-defined personas for per-call model routing, open since 2026-08-15; triage comment on [#102791](https://github.com/NousResearch/hermes-agent/pull/102791), 2026-09-04.
- OpenAI status, [Elevated errors across ChatGPT and Codex](https://status.openai.com/incidents/01M1KWEDH417T2CF44YYHZDFCR), 3 September 2026, read 2026-10-17.
- Claude status, [Elevated errors for multiple models](https://status.claude.com/incidents/461yvfrzpwtt) and [Elevated errors for Claude Sonnet 5](https://status.claude.com/incidents/288w7p4hk1l1), 3 September 2026, read 2026-10-17.