Choosing hardware for AI coding agents and 24/7 automation
Agents · AI Node Works engineering team · Published
The short answer: First decide where the model runs. If your agents call a cloud model API, the machine only runs tools, code, browsers and workflows: a reliable CPU system with plenty of memory and fast storage is enough, and no large GPU is needed. If the model runs locally, size the GPU for the model, the long context agents use, and how many agents run at once. Either way, build for unattended operation: ECC memory, cooling for sustained load, a UPS, monitoring and secure remote access.
Two kinds of agent machine
| Agents call a cloud model | Agents use a local model | |
|---|---|---|
| GPU | Not needed for the model | Sized to the model, context and number of agents |
| Processor and memory | Enough cores and RAM for builds, tests, browsers and containers | As left, plus memory to feed the GPU |
| Storage | Fast SSD for repositories, containers and caches | As left, plus space for model files |
| Depends on the internet | Yes, for every model call | Only for updates and outside tools |
| Code and data leave the network | Yes, in every prompt sent to the provider | Only where agents call outside services |
Many teams start with the first and move to the second when usage grows or when code and documents must stay inside. The machine for the first is modest; the second is a GPU system that also has to behave like a server.
Sizing a local model for agents
Agents differ from chat in two ways that both raise memory. They carry long context: source files, tool output and instructions add up quickly. And they run in parallel, several tasks at once or several agents cooperating, and each running task needs its own KV cache.
Runtimes make this explicit. Ollama's documentation notes that serving requests in parallel multiplies the context memory required by the number of parallel requests. Serving runtimes such as vLLM batch requests continuously and allocate cache as each conversation grows, which uses memory more efficiently but does not remove the peak.
| Agents at once | KV cache | Total estimate | Smallest launch configuration that fits |
|---|---|---|---|
| 1 | 3.2 GB | 23.6 GB | 1× AMD Radeon AI PRO R9700 |
| 5 | 16.1 GB | 37.0 GB | 2× AMD Radeon AI PRO R9700 |
| 10 | 32.2 GB | 53.7 GB | 2× AMD Radeon AI PRO R9700 |
Memory estimates from the site's VRAM calculator, assuming each agent reaches full context at the same moment. They say nothing about speed.
Speed matters more for agents than for chat, because an agent may make dozens of model calls to finish one task. Long prompts make prompt processing, not just text generation, a large share of the wait. Choose models and hardware with that in mind, and test with your own tasks.
What makes a machine reliable enough to run unattended
- ECC memory, so single-bit memory errors are corrected rather than silently corrupting a long job. Check it per component: AMD, for example, lists ECC on the Radeon AI PRO R9700 as supported on Linux only.
- A power supply with headroom for sustained full load, not just peak figures.
- Cooling sized for hours at full load, in a room that stays cool overnight.
- A UPS, and firmware set to power the machine back on after an outage.
- Remote access that works when you are not in the office, through a secured route, not an open port.
- Monitoring and alerts for GPU temperature, disk space and failed services, and services that restart themselves.
- Backups of configuration, prompts and any data the agents produce.
Agents run code, so isolate them
A coding agent writes and executes code, installs packages and calls tools. Treat it like a new team member with no track record.
- Run agents in containers or virtual machines, not directly on the host.
- Give them their own credentials with the least access that works, and never production credentials by default.
- Put the machine on a network segment that limits what it can reach.
- Keep logs of what agents ran and changed, and review them.
Questions to answer before buying
- Will the model run locally, in the cloud, or both?
- Which models, and how much context do your tasks really use?
- How many agent tasks run at the same time at peak?
- Must code and documents stay on your network?
- Where will the machine live, and who looks after it when it stops?
Sources Last verified 27 Oct 2026