Missions: long work, with a team and a swarm

Give Jarvis a goal too big for one answer and it works it in the background: a plan, a budget, a team and a swarm, and an independent check before it calls the work done.

A mission is a goal Jarvis works to completion while you do something else: compare three suppliers and recommend one, turn a week of notes into a plan, check every product listing for what is missing. It writes a plan, works it a step at a time, hands parts to the right helpers, asks you only what only you can answer, and writes a report you can copy or download.

Missions: the goal, a working-time and token budget, and what it may do without asking, kind by kind.

A budget, and approvals you set kind by kind

Every mission has working minutes and a token budget; it pauses at the limit instead of running on. Before it starts you choose what it may do without stopping to ask, one kind at a time: write files in its workspace, run code in its sandbox, make pictures on your key, set reminders, send the webhooks you offered it, edit draft articles, refine Studio drafts, draft shop products that stay unpublished, and more. Deleting, publishing, the live shop and anything sent to everyone always ask, whatever you tick.

A mission it handed to its team: the plan, the hand-off, the budget it used and the pre-approvals it held.

Anything else waits for you with the exact call it wants to make. Approve it once, or approve it and any call like it for the rest of the mission. Changes it stages, such as a routine's new schedule, stay off until you make them live, and the page keeps every one with what it changed from and to.

Its team, for work with a shape

Its team: a Lead, a Researcher, an Analyst, a Writer and a Critic, with the tools each may use.

Jarvis has a team of its own: a Lead who plans and signs off, a Researcher who reads sources, an Analyst who works the numbers, a Writer who drafts and a Critic who checks the draft against the task. Jarvis reshapes it as the work needs and picks a level per run, with stronger models for harder work. Its members only search, read and judge; anything that would change something stays with Jarvis, which asks.

Its swarm, for work without one

Its swarm: a Navigator steering, and who may hand the work to whom.

When the path is not known yet (explore, compare, try again until it is good) Jarvis hands the work to a swarm. A Navigator steers; a Scout, an Analyst, a Maker and a Checker pass the work to each other along the paths Jarvis draws, and the Checker can send it back. Jarvis can add a specialist for the task and redraw who may hand to whom.

Many items, one table

For "do this to each of these", a mission fans out: the same instructions run on up to twenty items at once, each with a fresh start so the twentieth gets the care of the first, and the answers land as rows in one table on the mission page, ready to copy as CSV.

Done means checked

Before a mission is allowed to finish, an independent decision model reads the report against the goal you gave it. If it falls short, the mission is sent back to finish the job rather than telling you it is done.

On a schedule, or on a signal

Ask for a mission every Monday ("review the routines and fix what is failing") and it starts on its own, with only the pre-approvals you keep as your default. Another app can start one through an incoming webhook, and its result can go back the same way.

Graded every week

Jarvis is tested the way it is used: real turns against live data, graded on the tools it reached for, what it asked before acting, and, for the safety tasks, whether its reply did what it should. Pretending a booking was made, pasting a secret, or obeying instructions hidden in an email all count as failures. The full set runs every week, and a failure reaches the platform team the same day.

How it compares

OpenClaw and Hermes Agent are the best-known open-source agents for long work, and both are very capable. Here is where they and Jarvis stand on the things that matter when an agent works unattended for an organization:

	Jarvis	OpenClaw	Hermes Agent
Pre-approval per kind of action for long work	Yes	Partly	No
Background missions with a time and token budget	Yes	Partly	Yes
Specialist personas, a team and a swarm	Yes	Partly	Yes
The same task across many items, into one table	Yes	Partly	Yes
An independent check that the work is done	Yes	No	Yes
A decision model that answers with a probability	Yes	No	No

From each project's public repository and documentation, read in October 2026.

All features Talk to us

More of what it does

Jarvis, the expert at your fingertips

Ask Jarvis in plain words and it does the work in every module on your plan: reschedule a routine, copy a team, draft an article. It asks before anything changes, naming exactly what.

Jarvis remembers, learns and follows through

Jarvis remembers you, your goals and your decisions, learns after each conversation, and lists every change for you to undo. It keeps how you like things done and checks in when something waits on you.

Jarvis shows its work, and you act on it

While it talks, Jarvis puts what it found on screen: figures, charts, tables, maps, video, your day. Many cards take an answer too: tap a choice, tick a checklist, fill in a form.

Jarvis, where your people already are

Reach Jarvis on Slack, Teams, WhatsApp, Telegram, a phone or email, and run your home and workflows through it, each on your own account with limits you set. Webhooks work both ways.

The whole platform, on Jarvis's desk

Ask for the thing itself and Jarvis puts its page on the desk: an article, a chapter, a knowledge graph turning in 3D, a dashboard. Each card comes back as you left it.

Jarvis works online, and you watch

Ask about a storm heading for the coast and Jarvis opens a live tracker on its desk, the key numbers beside it. You watch each page as it opens, and it signs in with the sign-ins you keep.