Agent comparisons
Gemini Spark vs Hermes Agent: Workspace or Your Runtime
Choose Gemini Spark when the work sits in Gmail, Calendar, Drive, and Docs and you want Google to run that agent. Choose Hermes Agent when you want to pick the model and keep the task instructions in files you can read. A managed Hermes instance stays online after you close your laptop. It does not connect those Google apps, and a Google model key does not change that.
Gemini Spark is the personal agent in the Gemini app. It is not the Gemini chatbot, and it is not a Gemini model you call from an API. Hermes Agent is Nous Research's open-source runtime, not the Hermes language-model family. You can point Hermes at another provider entirely.
How do Gemini Spark and Hermes Agent compare?
| Decision | Gemini Spark | Managed Hermes |
|---|---|---|
| What you start with | An agent inside the Gemini app | An instance you configure |
| Connected Google apps | Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube, and Maps, off by default | No native Workspace connectors |
| Unattended time | Schedules: time-based or conditional triggers | No managed cron on Hermes |
| Laptop closed | Google's cloud agent keeps working | The hosted instance keeps running |
| Browser | Chrome auto browse can use logged-in accounts and saved passwords, with permission | Optional Hosted Browser; you sign into a shared profile |
| Model | Google's Gemini models, including Gemini 3.5 on the May 19, 2026 description | A provider you connect, including Google's API if you choose it |
| Before a risky action | Designed to check with you before major actions, including spend and send | Messaging allowlists are not a per-action approval |
| Price | A Google AI subscription; no tariff on the Spark page | Hosting from $6.90/month; model access separate |
The Spark column follows the Gemini Spark product page reviewed on September 24, 2026, plus the Google posts cited below. The Hermes column is Lobsterland's managed Hermes hosting, not every option in the upstream project.
When is Gemini Spark the better fit?
Spark's own examples are Workspace jobs: a Monday inbox recap that also creates calendar blocks, a skill built from the last 50 emails you wrote, a Drive scan turned into a spreadsheet, a new client email that creates a Sheet row and a Drive folder. If that is the job, Spark is the product built for it. Google says you do not need to code, connections start off, and you turn the agent on yourself.
Hermes is the stronger candidate when the thing you want to own is the method: which model answers, which instructions it must follow, and which accounts it may touch. Hosting removes the server. It does not install Gmail. If the value you want is "read my inbox and write my calendar," configuring an open runtime is the long way around, and it still will not arrive at Spark's connectors.
Google's May 19, 2026 introduction says Spark runs on Gemini 3.5, uses the Antigravity harness, and is a cloud agent that continues after the laptop closes or the phone locks. That is a reason to pick Spark when you want Google's stack end to end. It is not a reason to treat every always-on agent as interchangeable with it.
Does managed Hermes keep running when your laptop is off?
Yes. The instance runs on hosting infrastructure, so closing your laptop does not stop it. Spark's product page makes the same kind of promise for Google's cloud agent: it can work in the background even when your phone and laptop are off. The shared property is that the process is not your laptop. The access is not shared.
A Hermes instance that is merely online can answer in built-in chat, Telegram, or Slack. It cannot scan Gmail because the laptop is closed. Nothing in the hosted environment is signed into your Google account unless you sign a browser session in yourself, and that session is the hosted browser described next, not your laptop's Chrome profile.
How does browser access differ?
On July 30, 2026, Google described Chrome auto browse for Spark. With your permission, Spark can use logged-in accounts and saved passwords for web errands, such as scheduling apartment viewings or researching flights and starting a booking. Google says sensitive actions such as payments are handed back to you, and that this Chrome capability was initially rolling out in the United States.
Managed Hermes does not attach to that Chrome profile. The optional Hosted Browser is a browser beside the instance. You open it, sign into the site the task needs, and the agent uses that same persistent profile. A login on your laptop does not appear there. Saved passwords from your own Chrome are not imported. There is no product step that hands a payment back to you; if a site asks for payment, you are in that hosted profile yourself, and you should not leave payment methods in a profile the agent can operate unless that task requires it.
The Spark page also says Gemini can research across sites and help complete bookings. We have not watched that flow. Do not read "both products mention a browser" as "both products have your Chrome passwords."
Does a Google model key replace Gemini Spark?
No. Lobsterland can connect Hermes to Google as a model provider, alongside Anthropic, OpenAI, OpenRouter, and MiniMax. The provider sends the prompt to that API and returns a completion. Spark's connectors are a separate switch, off by default, for Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube, and Google Maps.
Using a Gemini model inside Hermes is closer to OpenClaw's Gemini provider setup than to Spark: you are choosing an inference vendor. You are not granting the agent your Drive. If the task needs a specific Google file, paste the excerpt or export it. Do not assume the model name "Gemini" unlocked the apps.
Google's Spark FAQ says the agent does not read email indiscriminately and works under your direction. That is a product statement about Spark. It does not describe what a Google API key does inside Hermes, and it is not a privacy audit. We did not review a Spark training policy for this comparison.
What do schedules, skills, and confirmations cover?
Spark groups the work into Tasks, Skills, and Schedules. A Schedule is a time-based or conditional trigger. The Monday inbox example on the product page is that feature: scan mail, recap it, and block time on the calendar. Skills, in Spark's example, can be built in conversation from mail you have already written.
Hermes skills on Lobsterland are files you can install and inspect. The Hermes skills guide is the hosted path. They are not a button that reads your last 50 sent emails. Managed Hermes also has no cron surface, so it will not fire that Monday job for you. Upstream Hermes documents more automation than we host. If an unattended schedule is the requirement, read OpenClaw vs Hermes before you pick a runtime. OpenClaw's scheduler still does not add Gmail.
On confirmations, the May 19, 2026 post says Spark is designed to ask before high-stakes actions such as spending money or sending email. The product page says it checks before major actions, and the page footer tells you to supervise and interrupt. We have not verified those prompts in an account. Hermes allowlists decide who may message the agent on Telegram or Slack, and a group @mention decides when it should reply. Those controls do not require the agent to stop before every tool action. If a workflow must stop before it sends, do not treat an allowlist as that stop.
What do access and cost look like?
As of the product page reviewed on September 24, 2026, Spark is for Google AI Pro and Ultra subscribers, people 18 and older, in select countries, and for select business users. Google says access is still expanding. A July 30, 2026 post said Pro access was starting to roll out in more than 160 additional countries that day. The later product page still says "select countries" and names both Pro and Ultra. Use the account screen, not either sentence alone, to see whether you have Spark. The product page requires a subscription and does not print a price, so this comparison does not call either option cheaper.
Lobsterland's Hermes hosting starts at $6.90 per month. Every plan includes a seven-day trial: add a card up front, and the hosting charge happens when the trial ends unless you cancel first. Model access and optional add-ons are separate. The hosting pricing page lists the capacity tiers. A $6.90 instance plus a Google API bill is not a substitute price for a Google AI plan, because it does not include Spark.
How should you test Hermes against a Workspace-shaped task?
Use a task that does not pretend the agent has your inbox. Paste a redacted thread you are willing to share, then ask: "List the decisions and the open questions. Quote the line each one came from. Draft a reply I can edit. Do not send anything." This is an evaluation prompt, not a measured result.
- Launch Hermes and connect one model. Run the prompt in built-in chat. If you pick Google here, you are only choosing the model API.
- Check the quotes. Every decision should point at a line you pasted. A confident summary with no quote is a failed check, not a finished assistant.
- Save the instructions. When the shape is right, turn it into a skill and read the file. The next run should follow that file, not a hidden connector.
- Add one extra surface only if the task needs it. Telegram or Slack changes where you talk to it. Hosted Browser changes which website session it can see. Neither one reads Gmail.
Keep the rest of the surface honest. Managed Hermes includes skills, logs, upgrades, Paperclip, and optional Hosted Browser. The native dashboard is available from image v2026.6.5. It does not include managed cron, a workspace-file editor, multi-agent routing, usage analytics, or WhatsApp. Spark wins when those Google apps and schedules are the product you want. Hermes wins when you would rather configure the model and the method than hand Google the apps.
People comparing other personal agents can read Instinct vs Hermes Agent. That product is a different assistant, not a Workspace connector.
Frequently asked questions
Is Gemini Spark the same thing as Gemini?
No. Gemini Spark is Google's personal agent inside the Gemini app. Gemini is also the name of Google's chatbot and of its model family. This page compares the Spark agent with Hermes Agent, not a Gemini model with a Hermes model.
Can a Google API key make Hermes read Gmail?
No. On Lobsterland, Google is one supported model provider. That connection sends prompts to Google's model API. It does not turn on Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube, or Maps.
Does managed Hermes keep working when my laptop is off?
The hosted instance keeps running after you close your laptop. That does not give it Gemini Spark's Workspace connectors or Chrome saved-password browsing. A task still needs the tools you actually configured.
Does Gemini Spark require a paid plan?
The Spark product page reviewed on September 24, 2026 says it is available to Google AI Pro and Ultra subscribers, people 18 and older, in select countries, and to select business users. Confirm the plan and country on your account. The page does not list a price.