Agents
If you already run a coding CLI in a terminal, you already know most of what a NEXUS agent is. The delta is small but it changes everything: an agent is a terminal session that keeps its name and instructions between launches.
A plain CLI session is anonymous and mortal. You open a terminal, paste or retype the context that makes the session useful — what project this is, what conventions to follow, what it should never touch — and when the window closes, all of that evaporates. The next session starts from zero, and so do you.
A NEXUS agent wraps that same CLI session in a persistent identity. It has a name you chose, a persona and standing instructions that load on every launch, a provider, and a working directory it returns to. Quit the app, restart the machine, launch the agent again — it is still the same colleague, not a stranger with the same tools.
What problem it solves
The real cost of CLI-driven work is not typing commands; it is re-establishing context. Every fresh session pays a warm-up tax: re-explaining the project, re-stating the rules, re-earning trust in the output. When you run one session a day, the tax is annoying. When you run five specialized sessions — one for reviews, one for a particular repository, one for research — it becomes the dominant cost, and it quietly pressures you into using one undifferentiated session for everything.
Naming and persistence dissolve that pressure. Because identity is durable, specialization becomes cheap: an agent can afford to be narrow, because being narrow no longer means being re-configured by hand every morning.
What it replaces
Agents replace the folk infrastructure people build around CLIs: shell aliases that launch a tool with a particular flag soup, prompt preambles kept in a notes app and pasted at session start, per-project instruction files that only work if you remember to be in the right directory. All of that becomes properties of a named thing that NEXUS launches for you.
Persistence is also the foundation the rest of NEXUS stands on. Memory only makes sense for a session that will exist tomorrow; the mesh only makes sense between parties that have stable names; teams only make sense when members are durable.
Honest limits
An agent is still a local process. It runs when your machine runs it. There is no cloud copy of the session, no server-side continuation while your laptop is closed — NEXUS deliberately does not sync terminals anywhere. What persists is the agent's identity and what it has recorded, not a live process.
An agent can only touch what it has been granted. Capabilities are explicit grants, not defaults. A peer that lists your agent may see it at a reference tier without being able to read from it or deliver work to it, and the withheld capability comes with a stated reason rather than a silent failure.
A name is not competence. Persistence makes context cheap; it does not make the underlying model better. An agent with wrong standing instructions is wrong consistently, which is precisely why its memory is inspectable on the Neural Map.
To create and configure agents, see the Agents guide.