Manifesto
Agents made the code cheap. They made your attention expensive.
Saggar is a native Mac app for supervising several coding agents at once. It exists because of three beliefs that sit slightly awkwardly together, and one experience that anyone running more than two agents already recognizes.
01
The agents work.
Saggar didn't come out of a view about where this technology is going. It came out of a Tuesday with four agents running, one of them stopped at a permission prompt, and no way to tell which one.
The argument about whether any of this is real is mostly being had by people who haven't left a coding agent running for a week. The shape of a working day has already changed. More of it is reading, deciding, and redirecting; less of it is typing. That is a genuine shift, and treating it as a fashion is a way of losing several years.
It is also uneven. An agent will produce a confident, plausible, wrong answer and hand it to you with the same tone it uses for a good one. A good day with agents beats a good day without them. A bad one is worse, and the bad ones are still frequent enough to plan around.
02
That makes the person at the desk matter more, not less.
The cheaper it becomes to produce code, the more expensive it becomes to decide which code should exist. Generation is abundant now. Judgement isn't. Neither is taste, architecture, knowing what not to build, or being answerable to a colleague for something that shipped.
An agent will do what you ask. It will not tell you the ask was wrong. Somebody still holds the intent, and holding intent across five parallel sessions is real work — the kind that doesn't look like work from the outside, because the screen is mostly other people's output.
So Saggar is built for someone who wants to stay in the loop rather than escape it. It doesn't try to run your agents for you. It tries to make supervising six of them cost about what supervising one used to.
This is a bet, not a proof. If judgement turns out to be as automatable as autocomplete was, we will have built a careful tool for a job that no longer exists. We don't think that's where this goes, but we're not going to pretend the question is settled.
03
We are not bullish on the business of AI.
The technology is real. Much of the industry around it is a land grab. Enormous money is chasing the model layer, and a tier of products above it is a thin coat of paint on somebody else's API with a valuation attached. Prices will move. Models will be deprecated. A fair number of the companies selling you an agent this year will not be around to support it next year.
We treat that as a design constraint rather than a complaint. Saggar is a native Mac app that runs on your machine and holds no opinion about which agent wins — six providers are entries in a catalogue, and adding the seventh is a small change. Your project's setup lives in plain files inside your checkout, readable and committable. The app is free and asks for no account.
If the model you rely on today is gone in a year, Saggar keeps working. If we are gone in a year, you still have your files, your terminals, and your Mac.
We sell something too. Remote control costs £1.99 a month, because reaching your Mac from a train genuinely needs a server and servers cost money. Everything that runs on your own machine is free, and we intend to keep it that way.
What it actually feels like
Running one agent is a conversation. Running six is a shift on a factory floor.
One is halfway through a refactor. One has been sitting at a yes-or-no prompt for eleven minutes. One finished, and its green test run is scrolling past unread. One failed in a way that still looks like working. You are holding all of that, in your head, alongside the problem you actually sat down to solve.
It produces a particular kind of tiredness — not the tiredness of a hard problem, but of held state. Plates, spinning, none of them individually difficult. From the outside it does not look like difficulty at all, which is part of why so little has been built for it.
That is the problem Saggar takes seriously. Not making you faster. Making supervision cost less.
What follows from it
Beliefs are cheap. These are the decisions they cost us, and you can check every one of them against the app.
- A session says what it needs
- Needs you, working, idle, finished, failed. Five words, decided from what the terminal is actually doing, so the question "which one stopped?" has an answer you can read rather than one you have to reconstruct.
- Working agents are left alone
- Progress is not news. A session that is getting on with it moves out of the way, and only a decision, a failure, or finished work is allowed to reach for you.
- One queue, not eleven tabs
- Everything waiting on a human forms a single ordered line. ⌘J walks it. The alternative is holding the queue in your head, which is exactly the part that wears people down.
- Quiet by default
- A tool that pings for everything trains you to ignore it, and then it is worse than nothing. Saggar interrupts for decisions and stays silent for activity.
- Your state stays yours
- Projects, commands, and session layout are plain files in your checkout and your home directory, readable without us. The Mac app needs no account, and nothing you run in it is stored anywhere but your own machine.
- You are allowed to leave
- Stepping away shouldn't mean abandoning six running sessions. Pair a phone, answer the one prompt that needs you from wherever you are, and close the lid on the rest.
Where we might be wrong
Saggar is a Mac app, so it is no use to you on Linux or Windows. It is not open source. When the lid closes, your agents stop. Free tools do some of this, and one of them may be the better answer for you — the comparison pages say where, in more detail than is comfortable.
The next decade of software gets written by people who are good at deciding, working with machines that are good at doing. That is a fair trade, as long as the deciding does not drown you. Saggar is one small argument about how to keep it from drowning you.