AI agents do things for you, semi-autonomously, and one question is how we coordinate with them.
By do things for you
I mean
(btw I use the heck out of Claude Code, despite there being better pure coding models available, proving that the difference is in the quality of the agent harness,_ i.e. how it approaches problems, and Anthropic has nailed that.)
By coordinate
what I mean is: once you’ve stated your intent and the agent is doing what you mean (2025), or it’s listened to you and made a suggestion, and it has actioned the tasks for that intent then how do you
- have visibility
- repair misunderstandings
- jump in when you’re needed
- etc.
Hey did you know that
15 billion hours of time is spent every year by UK citizens dealing with administration in their personal lives.
Such as: private bills, pensions, debt, services, savings and investments
(and also public services like healthcare and taxes).
It’s called the time tax.
I was chatting with someone last week who has a small collection of home-brew agents to do things like translating medical letters into plain language, and monitoring comms from his kid’s school.
It feels like small beans, agents that do this kind of admin, but it adds up.
Every agent could have its own bespoke interface, all isolated in their own windows, but long-term that doesn’t seem likely.
See, agents have common interface requirements. Apps need buttons and lists and notifications; agents need… what?
What is particular to agents is that they need progress bars not notifications (2023): after decades of making human-computer interaction almost instantaneous, suddenly we have long-running processes again. Thinking…
Agents sequence tasks into steps, and they pause on some steps where clarification is needed or, for trust reasons, you want a human in the loop: like, to approve an email which will be sent in the user’s name, or a make a payment over a certain threshold, or simply read files in a directory that hasn’t been read before.
Claude Code has nailed how this works re: trust and approvals. Let’s say you’re approving a file edit operation. The permission is cleverly scoped to this time only, this session, or forever; and to individual files and directories.
Claude Code has also nailed plans, an emerging pattern for reliability and tracking progress: structured lists of tasks as text files.
Ahead of doing the work, Claude creates a detailed plan and stores it in one of its internal directories.
You can already entice Claude to make these plans more visible - that’s what I do, structuring the work into phases and testing for each phase - and there’s discussion about making this a built-in behaviour.
Want to see a sample plan file? It’s just some context and a list of to-dos. Check out My Claude Code Workflow And Personal Tips by Zhu Liang.
So… if agents all use plans, put all those plans in one place?
Another quality that is particular to agents is that when you’re running multiple agents each running down its personal plan, and you have a bunch of windows open and they’re all asking for permissions or clarifications or next instructions, and it feels like plate spinning and it is a ton of fun.
Task management software is a great way to interact with many plans at once.
Visually, think of a kanban board: columns that show tasks that are upcoming, in progress, for review and done (and the tasks can have subtasks).
Last week on X, Geoffrey Litt (now at Notion) showed a kanban board for managing coding agents: When an agent needs your input, it turns the task red to alert you that it’s blocked!
There’s something in the air. Weft (open source, self-hosted) is
a personal task board where AI agents work on your tasks. Create a task, assign it to an agent, and it gets to work. Agents can read your emails, draft responses, update spreadsheets, create PRs, and write code.
It is wild. Write something like Create a cleaned up Google Doc with notes from yesterday’s standup and then send me an email with the doc link
and then an agent will write actual code to make the doc, summarise the notes, connect to your Gmail etc.
This is great in that you can instruct and then track progress in the same place, you can run many tasks simultaneously and async, and when you jump in to give an approach then you can immediately see all the relevant context.
Ok great self-driving to-do lists.
But wouldn’t it be great if all my agents used the same task manager?
Is it really worth special-casing the AI agent here?
Linear is a work to-do list. Sorry, a team collaboration tool oriented around tickets.
Linear for Agents is smart in that they didn’t launch any agents themselves, they simply built hooks to allow AI agents to appear like other users, i.e. the agent has an avatar; you can tag it etc:
Agents are full members of your Linear workspace. You can assign them to issues, add them to projects, or @mention them in comment threads.
(Agents are best seen as teammates (2023).)
In the general case what we’re talking about is a multiplayer to-do list which AI agents can use too.
Really this is just the Reminders app on my iPhone?
Long term, long term, the Schelling point for me, my family, and future AI agents is a task manager with well-scoped, shared to-do lists that I already look at every day.
Apple is incredibly well placed here.
Not only do they have access to all my personal context on the phone, but it turns out they have a great coordination surface too.
So Apple should extend Reminders to work with agents, Linear for Agents style. Let any agent ask for permission to read and write to a list. Let agents pick up tasks; let them add sub-tasks and show when something is blocked; let me delegate tasks to my installed agents.
Then add a new marketplace tab to discover (and pay for) other agents to, I don’t know, plan a wedding, figure out my savings, help with meal planning, chip away at some of that billions of hours of time tax.
The Reminders app is a powerful and emerging app runtime (2021) – if Apple choose to grab the opportunity.
AI agents do things for you, semi-autonomously, and one question is how we coordinate with them.
By I mean
(btw I use the heck out of Claude Code, despite there being better pure coding models available, proving that the difference is in the quality of the agent harness,_ i.e. how it approaches problems, and Anthropic has nailed that.)
By what I mean is: once you’ve stated your intent and the agent is doing what you mean (2025), or it’s listened to you and made a suggestion, and it has actioned the tasks for that intent then how do you
Hey did you know that
Such as: (and also public services like healthcare and taxes).
It’s called the time tax.
I was chatting with someone last week who has a small collection of home-brew agents to do things like translating medical letters into plain language, and monitoring comms from his kid’s school.
It feels like small beans, agents that do this kind of admin, but it adds up.
Every agent could have its own bespoke interface, all isolated in their own windows, but long-term that doesn’t seem likely.
See, agents have common interface requirements. Apps need buttons and lists and notifications; agents need… what?
What is particular to agents is that they need progress bars not notifications (2023): after decades of making human-computer interaction almost instantaneous, suddenly we have long-running processes again.
Agents sequence tasks into steps, and they pause on some steps where clarification is needed or, for trust reasons, you want a human in the loop: like, to approve an email which will be sent in the user’s name, or a make a payment over a certain threshold, or simply read files in a directory that hasn’t been read before.
Claude Code has nailed how this works re: trust and approvals. Let’s say you’re approving a file edit operation. The permission is cleverly scoped to this time only, this session, or forever; and to individual files and directories.
Claude Code has also nailed plans, an emerging pattern for reliability and tracking progress: structured lists of tasks as text files.
Ahead of doing the work, Claude creates a detailed plan and stores it in one of its internal directories.
You can already entice Claude to make these plans more visible - that’s what I do, structuring the work into phases and testing for each phase - and there’s discussion about making this a built-in behaviour.
Want to see a sample plan file? It’s just some context and a list of to-dos. Check out My Claude Code Workflow And Personal Tips by Zhu Liang.
So… if agents all use plans, put all those plans in one place?
Another quality that is particular to agents is that when you’re running multiple agents each running down its personal plan, and you have a bunch of windows open and they’re all asking for permissions or clarifications or next instructions, and it feels like plate spinning and it is a ton of fun.
Task management software is a great way to interact with many plans at once.
Visually, think of a kanban board: columns that show tasks that are upcoming, in progress, for review and done (and the tasks can have subtasks).
Last week on X, Geoffrey Litt (now at Notion) showed a kanban board for managing coding agents:
There’s something in the air. Weft (open source, self-hosted) is
It is wild. Write something like and then an agent will write actual code to make the doc, summarise the notes, connect to your Gmail etc.
This is great in that you can instruct and then track progress in the same place, you can run many tasks simultaneously and async, and when you jump in to give an approach then you can immediately see all the relevant context.
Ok great self-driving to-do lists.
But wouldn’t it be great if all my agents used the same task manager?
Is it really worth special-casing the AI agent here?
Linear is a work to-do list. Sorry, a team collaboration tool oriented around tickets.
Linear for Agents is smart in that they didn’t launch any agents themselves, they simply built hooks to allow AI agents to appear like other users, i.e. the agent has an avatar; you can tag it etc:
(Agents are best seen as teammates (2023).)
In the general case what we’re talking about is a multiplayer to-do list which AI agents can use too.
Really this is just the Reminders app on my iPhone?
Long term, long term, the Schelling point for me, my family, and future AI agents is a task manager with well-scoped, shared to-do lists that I already look at every day.
Apple is incredibly well placed here.
Not only do they have access to all my personal context on the phone, but it turns out they have a great coordination surface too.
So Apple should extend Reminders to work with agents, Linear for Agents style. Let any agent ask for permission to read and write to a list. Let agents pick up tasks; let them add sub-tasks and show when something is blocked; let me delegate tasks to my installed agents.
Then add a new marketplace tab to discover (and pay for) other agents to, I don’t know, plan a wedding, figure out my savings, help with meal planning, chip away at some of that billions of hours of time tax.
The Reminders app is a powerful and emerging app runtime (2021) – if Apple choose to grab the opportunity.