OmniscioGuidesAI agent dashboard
What has to be on an AI agent dashboard?
An AI agent dashboard is one screen that answers three questions about every agent you have running: what is it doing, which one is waiting on you, and what is it costing you. If answering any of those means opening a transcript, you have a session list rather than a dashboard. This guide covers what each of those three jobs takes to do properly, six things to check before you put a screen in front of your agents, and what the option built into Claude Code covers.
- API refactorNeeds You
- Dependency upgradeRunning
- Test suiteRunning
- Docs syncReady
By the Omniscio team. Updated .
How this page was checked
One product other than ours is described in its own right here: Anthropic's agent view, read from Anthropic's own agent-view page on 22 October 2026, including the limitations section at the foot of it. Every line about it rests on a dated row in this page's fact ledger, which sits beside the page source and ships nowhere.
No dashboard named anywhere on this page was installed or run for it, and no search-result snapshot was taken. Where a page says what a product does, it is repeating that product's own documentation and saying which day it was read. The one thing that could not be checked from the machine this was written on is who currently ranks for these searches, so this page makes no claim about that.
What is an AI agent dashboard for?
One agent in a terminal is easy to supervise: you watch it. Several agents at once change what the watching is made of. The output scrolling past is not the problem. The state is. Which one is still working, which one stopped to ask you something twenty minutes ago, which one finished and is now sitting on a branch nobody has merged, which one has been retrying the same failing command since you left.
Those are not questions a terminal answers, because a terminal shows you whichever session it is attached to. So a dashboard is not a bigger terminal. It does three jobs: show every agent's state in one view, put the agents that need a person where you cannot miss them, and let you answer them without attaching to anything.
That third job is what separates a dashboard from a watch window. If a one-word reply costs you a context switch into a full transcript and back, the screen is showing you the work rather than letting you do it.
What has to be on one?
Six things, in the order that a missing one hurts. The third column is the part worth reading: it is what the same feature looks like when it is only half done, and it is the half-done version most screens ship.
| Job | What it looks like when it works | What tells you it does not |
|---|---|---|
| Every agent's state, in one word | A fixed vocabulary, where the same condition always gets the same word, with colour doing a second pass over it. | A free-text last line. You read five rows of prose to work out which of them is waiting on you. |
| The ones that need you, at the top | Waiting agents sorted above working ones, and a row that starts waiting moves back up while you watch. | Your own memory as the sort order, and a scroll to check whether anything moved. |
| What each one is waiting for | The question, the permission or the plan approval it stopped on, on the row itself. | Attaching to the agent to find out what it asked. |
| A way to answer without attaching | Type the reply on the row and send it. The agent keeps its place in the list. | Opening the whole transcript, scrolling to the bottom and typing into the live session. |
| What it is costing | A running figure per agent, beside the agent, so a runaway loop is visible while it is running. | A monthly total you go and read afterwards, once the loop has finished spending. |
| What happened while you were away | A history of the agents you were not watching, readable or searchable after the fact. | Output that scrolled past a terminal buffer that no longer holds it. |
How do you tell a real one from a pretty list?
Six checks. Each takes a few minutes with any tool you are considering, and each one is hard to pass without having actually built the thing.
- Ask it what it knows about waiting. A screen that cannot separate a working agent from an agent stopped on a question has redrawn the problem rather than solved it. Start two agents, ask one to make a decision only you can make, and see whether the list changes shape.
- Try to reply from the list. If answering costs you a context switch, you will start postponing answers, and a postponed agent is a stopped agent.
- Watch what happens when the machine goes to sleep. Some screens are a view onto processes that die with the laptop lid. Others resume. Which one you have decides whether the screen is for this afternoon or for a long run.
- Look for the cost before you need it. A per-agent figure next to the agent is the difference between noticing a loop in ten minutes and noticing it at the end of the month.
- Read the state words. If a row can say anything, nothing on the row can be trusted at a glance, and you are back to reading prose.
- Check the order. Start a third agent halfway through and see whether a newly waiting one takes the top spot. Sorting that only happens when the list is rebuilt is the most common way a dashboard quietly stops being one.
What are the options today?
The built-in one is worth understanding first, because if it covers what you need you can stop reading. Anthropic ships agent view inside Claude Code, opened with claude agents or by pressing ← on an empty prompt. It is one screen for your background sessions, grouped into four: Ready for review and Needs input above Working and Completed, with a count at the top of how many are waiting, working and finished. You dispatch a new task from the input at the bottom of the same screen, press Space on a row to see its most recent output or the question it is waiting on and reply without leaving, and press Enter to attach when you want the full conversation. The sessions keep running with no terminal attached, so you can close the screen and come back to them.
What its own documentation lists as limits, and what its page does not describe:
- Sessions are local. They run on your machine, are preserved across sleep, and stop if the machine shuts down.
- Quota is spent per session. Background sessions draw on your subscription the way interactive ones do, so ten in parallel use it roughly ten times as fast as one.
- A worktree goes with its session. A worktree Claude Code created is deleted when the session is, so anything in it has to be committed first.
- It is a research preview. Anthropic says the interface and the keyboard shortcuts may change.
- One harness. Its page describes Claude Code sessions; it does not describe another harness's agents appearing in the same list.
- No per-agent spend figure. Its page describes session state, timing and answers, and does not describe a cost readout.
Third-party options are covered elsewhere on this site rather than repeated here: the guide to running 10 to 20 coding agents at once compares the harnesses and the tools people hold them in, and the comparison pages go through them one at a time. What this page adds is the question underneath all of them, which is what the screen itself has to do before the tool around it matters.
Where does Omniscio fit?
Omniscio is the app we make, and the six jobs above are what it was built around. It is worth saying plainly what it is: a desktop app that runs your agents in one window and keeps every session's state in a single list, rather than a replacement for any harness. It runs on Windows, macOS and Linux and drives the agents you already use.
| The job | What Omniscio shows |
|---|---|
| Every agent's state, in one word | Eleven fixed statuses, so the same condition always gets the same word, with colour doing the second pass: Running green, Needs You amber, Error and Stalled red, Waiting in ember orange for an agent parked by a capacity limit. |
| The ones that need you, at the top | Needs-you sessions sort first, and error or stalled ones group into an Interrupted section above the live ones rather than vanishing into the list. |
| What each one is waiting for | The attention state carries the reason, so the row says whether it is a question, a plan approval, a permission request, a rate limit or a suspension. |
| A way to answer without attaching | One inbox gathers what is waiting on you from every project, so answering is one pass rather than six open windows. Typing mid-turn reaches the live agent without losing its context, and interrupting it keeps the session alive for your next instruction. |
| What it is costing | A cost figure sits in each session's header and is marked as an estimate when a provider reports tokens rather than a true dollar amount, so a runaway loop is visible while it runs. |
| What happened while you were away | Full-text search across every conversation, and a desktop notification when a session moves into a state that needs you. |
- The same list on a phone browser, over a paired connection, for checking what a session is doing from somewhere else. Reaching it from the public internet is a switch that is off by default.
- The Dashboard, the eleven session statuses and the inbox are each documented page by page, so you can read exactly what a state means before you rely on its colour.
- Sorting and filtering a long list, notification rules and per-session cost are covered in the session list, notifications and cost control. Running several agents across projects at once is documented in full.
The features page lists what is in it today, and the download page has the builds.
Sources
Every factual line on this page rests on one of these, and the date beside each is the day it was read.
- Anthropic: Manage multiple agents with agent view, read 22 October 2026, for how agent view is opened, the four groups and the counts at the top, dispatching from the same screen, peek and reply, and the limitations section behind the local-sessions, quota, worktree and research-preview lines.
- What Omniscio does comes from our own pages, read 22 October 2026: the features page for the status vocabulary and its colours, the attention-first inbox, the per-session cost figure and the phone view, and the documentation library for how each piece behaves.
Anthropic's agent view is in research preview and its own page says the interface may change. Re-open that page before you trust this one on any detail of it.
Common questions
What is an AI agent dashboard?
One screen that answers three questions about every agent you have running: what is each one doing, which one is waiting on you, and what is it costing. A screen that shows output but not those three things is a session list with a nicer font.
Does Claude Code have a dashboard built in?
Yes. Running claude agents opens agent view, one screen for your background sessions grouped into four: Ready for review and Needs input above Working and Completed. You dispatch a new task from the input at the bottom of that screen and reply to a session without leaving it. Its own documentation lists three limits: sessions run on your machine and stop if it shuts down, ten sessions use subscription quota roughly ten times as fast as one, and a Claude-created worktree is deleted with its session.
Do I need a separate app to keep an eye on several agents?
Not to run them. Several terminals and a status convention will do it, and every criterion on this page can be met by hand. What a purpose-built screen buys you is the bookkeeping: one fixed vocabulary for state, waiting agents sorted to the top, and a cost figure you did not have to go and find.
How many agents should one dashboard hold?
The limit is your attention rather than the list. Past roughly five running at once, what matters is that the ones waiting on you are impossible to miss; a screen that sorts them to the top handles sixteen, and one that does not struggles with four.
See every agent on one screen
Install Omniscio, start the agents you were going to start anyway, and watch the ones that need you come to the top. The features page lists what is in it today, and the Dashboard documentation walks through the screen itself.