AI for Research Efficiency, episode 14: Where your agent lives
Narration: an AI-generated voice (ElevenLabs).

[0:00] Where your agent lives
Episode 4 said: leave it on. But by default, a plugged-in Windows 11 PC can sleep after five minutes.
And updates can restart it outside active hours: five in the evening to eight in the morning.
AI for Research Efficiency. Episode 14: Where your agent lives.
Last time, you saw how this series is made.
This time: where an agent can live, and how to keep it running and reach it safely.

[0:33] 1. The rungs
One: the rungs.
In a browser, the assistant works on copies you upload. An agent on your laptop works in your own folders, while the laptop is awake.
The owner of this series went further: first a small always-on computer, for long runs steered from somewhere else.
The owner's was a Beelink SER8, a mini PC with 16 gigabytes of memory, known on the network as newt.
Then a server with local tools, in the owner's words, to help ensure both strict confidentiality and cost control. Each rung adds something, and costs something.

[1:14] 2. Leave it on
Two: leave it on.
OpenAI's own documentation suggests a machine like this: use a dedicated always-on Mac or Windows PC when you want ChatGPT to stay reachable for longer-running work.
The checklist, Windows first. Sleep: when plugged in, set Make my device sleep after to Never; on a Mac, turn on Prevent automatic sleeping. And on a laptop, closing the lid can still put it to sleep.
Some agents keep it awake while they run: Antigravity, Claude Code on a Mac, and Codex with an experimental setting, off by default.
Updates: stretch active hours over your runs, up to 18 hours, and restart it yourself between runs. After a power cut, a Mac desktop can start up by itself, and so can some mini PCs.

[2:09] 3. Keep the session alive
Three: keep the session alive.
Close a terminal, and the agent in it stops; each tool has a way around that. Claude Code's background sessions need no terminal. They're a research preview, and a shutdown or restart stops them.
Antigravity's remote control can run as a service; on Windows, starting it with the computer needs an administrator.
Codex's always-on host is the ChatGPT desktop app. And on a Mac, Linux, or WSL, tmux keeps an agent running after you disconnect.
Cloud scheduled tasks, in Claude Code and Codex, run with your computer off, on paid plans. Anything that needs your own files needs the machine on.

[2:55] 4. Reach it from anywhere
Four: reach it from anywhere.
Each agent has a remote control. Claude Code's and Codex's need a paid plan; Antigravity's documentation names no plan.
For everything else, there's SSH, a terminal on the other machine, over Tailscale, a private network that works without opening ports on your router.
The vendors give two cautions: don't expose it directly to the internet, and sign in with keys, not passwords.
And Windows Remote Desktop needs Windows Pro on the PC you connect to.
The owner's own route changed over time: first Remote Desktop; then WireGuard, the protocol Tailscale is built on, and SSH; now an SSH app, with tmux in several panes.

[3:45] 5. The power bill
Five: the power bill.
In Arkansas, home power averaged about 13.7 cents a kilowatt-hour from January to August 2026.
Left on all year, 10 watts costs about $12, and 50 watts about $60. 300 watts costs about $360: some $30 a month, more than a $20 plan.
Apple lists four watts at idle for the Mac mini. For other machines, a plug-in meter settles it.

[4:19] 6. A server with local models
Six: a server with local models.
The owner's server runs its own language model, transcription and speech, on two large graphics cards.
It cost more than $30,000. By the owner's estimate, it has saved more than 360 hours to date, over roughly nine months, and made possible projects like video generation and testing; it's also used for editing and improving internal grant development work.
How good are local models? On one public index, in October 2026, the best closed model scored 58; the owner's model, 40; and models that fit one workstation, 12 to 34. Good enough, in our reading, for bulk reading, transcription and first drafts; not a replacement for the frontier on hard work.
Open weights don't share one license: Gemma 4's is Apache, a Qwen research license allows research or evaluation only, and Llama 4 asks you to display Built with Llama.
And local isn't the same as approved. At the University of Arkansas, any AI tool used on a university system or for university business must be vetted and approved.
NIH's ban on AI in review critiques names the technology, not where it runs, in our reading. Health data and controlled unclassified information bring rules for the machine itself.
One trap: in Ollama, a model whose name ends in :cloud runs on Ollama's servers.

[5:55] 7. Your agent and a local model
Seven: your agent and a local model.
Codex documents it: --oss, with Ollama or LM Studio. Claude Code works through a setting those two document, but loses its remote control.
Antigravity's command-line tool documents only Google's Gemini API; its Python SDK can run local models.
So, a pattern that keeps the data local: a frontier agent builds and tests the process on a made-up file with the same columns. Then you run it on the real data, with a local model.

[6:33] 8. Two ways not to buy
Eight: two ways not to buy.
Cloud agents keep working while your laptop sleeps: Claude Code's and Codex's on paid plans, and Google's Jules, free for 15 tasks a day.
But they work on a copy of your project, in the vendor's cloud, not on the files on your computer.
Or use your campus. Most of the Arkansas High Performance Computing Center's services are free to eligible researchers, with GPU jobs of up to 72 hours.
Nationally, NSF's ACCESS costs nothing, and you don't need an NSF award. The National AI Research Resource pilot requires open, publishable results.

[7:16] 9. Where should it live?
Nine: so, where should your agent live?
Runs that outlast the lid: a small always-on computer, or a cloud agent. Graphics cards now and then: your campus cluster or ACCESS, before you buy.
Confidential data: your campus's approved route first, and a local model only with that approval.
Wherever it lives, secure it: updates on your schedule, an encrypted disk, nothing open to the internet, and keys to sign in.

[7:48] Try it
Pause here, and try it: ask your agent to run the always-on check from the companion page. You should see a short report on sleep, updates and remote access, with nothing changed.

[8:02] Review and next
So: keep it awake, keep the session alive, reach it safely, count the power, and remember that local isn't the same as approved.
That's the route. Next: In Practice, starting with Ask for the process, not the answer.