➜ nextos / agents — one request, delegated to the specialists that own the tools.
Every app is an agent.
They work as a team.
Not one assistant with a hundred tools, a society of specialists. Every app carries an agent that owns its own tools, and the assistant is the conductor: it decomposes what you ask and hands each piece to the specialist that does it best, in parallel, pausing for your approval before anything irreversible.
- app agents
- 44
- tools they share
- 399
- model providers
- 5
- approvals you hold
- ∞
➜ “prep the Q3 launch: pull the notes, total the budget, draft the team mail, hold Thursday”
assistant · delegating to 4 specialists in parallel
- ✓app:notes find the Q3 launch notesdone
- ✓app:spreadsheet total the budget rowsdone
- ▸app:mail draft the note to the teamrunning
- ?app:calendar hold Thursday 2pmneeds you
approval required
app:calendar wants to create an event on your calendar. Allow it?
the conductor keeps the plan; the specialists do the reaching; you hold the approvals
01How it works
No agent to build.
The team is already there.
Every app ships an agent
Each app already exposes its tools. NextOS derives a specialist agent from them automatically, scoped to that app and named for it: a Notes agent, a Mail agent, a Terminal agent. No one authors forty agents by hand.
The assistant orchestrates
Type a sentence into the strip and the assistant plans the work, then delegates each self-contained piece to the specialist that owns it, fanning several out at once and bringing the answers back together.
You approve what matters
Every irreversible step pauses for your say-so, wherever in the chain it happens. You watch the whole run as a trace of tool calls, not a black box, and stop it at any point.
02Make it yours
Custom agents,
on a schedule.
When a job comes back often, give it a specialist of its own. Name it, write its brief, hand it the smallest tool set that does the work, and set it loose on a trigger. Chain agents into automation pipelines that run on a schedule, on an event, or on demand, and keep working on the server while the tab is closed.
- Custom agents inherit the workspace model unless you pin one
- A tight tool scope: an agent can never exceed the reach you grant it
- Automations run pipelines headless, with Web Push when they need you
- A sub-agent can never outreach the run that spawned it
A kernel that will not run away with your money
Every agent, built-in or yours, runs on one hardened loop. It is bounded on every axis that could otherwise spin, spend, or be manipulated.
- A per-tool timeout: a hung tool can never hang a run
- A wall-clock deadline and a stream-idle watchdog on every run
- Per-run token and cost caps that actually stop the run
- A no-progress guard that ends a loop instead of burning credit
- A trust boundary: tool output is data, never instructions
- A permission hook can only tighten a decision, never loosen a deny
no setup · the team is already assembled