AI Agents That Work While You Sleep
Moving beyond chat to background, ambient, and proactive AI
I’m Tanay Jaipuria, a partner at Wing and this is a weekly newsletter about the business of the technology industry. To receive Tanay’s Newsletter in your inbox, subscribe here for free:
Hi friends,
Most of us today use AI in a chat box. You type, it replies, everything happens in that one thread. That was the right starting point. It is the wrong place for a lot of the work we are now trying to hand off to agents.
Refactoring a codebase, watching the web for changes or building a briefing for tomorrow are not five second tasks. If you force those into chat and let it run for 10s of minutes, people inevitably get bored and leave.
For these tasks, The model should not be a slightly faster colleague who types back at you. It should feel more like a background process that you brief, that disappears for a while, and that hands you something useful where you actually work.
That is the shift I want to talk about: the rise of background agents, how they run, and where we already see them showing up.
What are background agents
A background agent is an agent you do not have to sit and watch. You might talk to it in chat, you might tag it in Slack, you might forward it an email, but once it understands the task, it runs elsewhere and comes back only when it has something worth your attention.
That already makes it different from a classic chatbot. In the chatbot world, every interaction is “you ask, it answers.” In the background world, the unit of interaction is a task. You brief the agent once, it may run for minutes or hours, and the result shows up as a pull request, a digest, a comment, or a notification.
On top of that, background agents can be:
Ambient: They are always running in some sense, and responding to some changes in some input. They wake up because the world changed, not just because you typed a prompt in. A new email arrives, a web page changes, a metric moves, a calendar event appears.
Proactive: They are allowed to tap you on the shoulder when there is something important, rather than waiting for you to ask at all. Think alerts, daily briefings, suggested decisions or actions (or at some point it may have taken these actions for you as well).
Not all background agents have to be proactive or ambient. You can have a very simple background agent that only acts when you tell it to. But once you accept that the work happens away from the chat window, it becomes natural to let agents respond to time and events as well as direct commands.
Chat is still useful. It is a great place to negotiate scope, explain edge cases, and set guardrails. It just does not have to be the only interface or the place where the heavy lifting happens.
What these agents actually look like
Broadly, background agents show up in two ways.
The first are the ones you explicitly create. Early on, that mostly happened in chat. You open something like ChatGPT and kick off a Deep Research request or a coding assistant, describe the task, and a background job spins up from that thread. Cursor or Codex style agents that take a spec, go off, and eventually return a pull request are good examples here.
What is changing is that these agents are now being created in the places where the work already lives. Instead of only briefing them in chat, you:
Assign an issue in Linear or Jira to an agent instead of a person
Tag an agent in a Slack or Teams thread and ask it to own that conversation
Forward an email to an agent address which can ingest the context and get started
You are still creating a discrete job, you are just doing it from tickets, threads, and emails rather than only inside a chatbot.
The second style is more ambient and always on. Here you create the concept once, not each individual run.
That might look like:
An agent that runs whenever new emails come into your inbox, monitoring your inbox and turning a pile of messages into a single summary and task list or drafting replies (e.g. Fyxer)
An agent that tracks changes on competitor sites or pricing pages and hands you updates so you can react (e.g. Web Monitoring Agents such as Yutori Scout)
An agent that watches or monitors a set of data for anomalies or changes and only surfaces something or takes action when it sees a meaningful pattern
Sometimes these run purely on events, sometimes they also run on a schedule, for example sweeping everything once a night and dropping a fresh brief in the morning (e.g. daily briefing agents that can be created on Lindy or Gumloop). Sometimes these can be derided as just cron jobs, but they also have LLMs baked in, and so are more powerful.
So you get two complementary patterns. Jobs you explicitly assign, often from the same tools you already use to assign work to humans. And agents that feel more like part of the environment, always running in the background, watching your inbox, data, or the web, and only appearing when there is something you might want to do about it.
Examples of background agents
To make this less abstract, here are three classes of background agents that already exist and feel meaningfully different from simple chatbots.
Coding agents as background teammates
Cursor is probably the cleanest example. You describe a change, sometimes in chat or directly assigning it in Linear. Cursor spins up cloud agents that read and edit your repo, run tests, and open pull requests. The work might run for quite a while. When you come back, the real interface is the diff and the test results.
Monitoring agents for the web
Products like Yutori’s Scouts or Exa-powered monitors allow you to create an ambient agent that continously monitors the internet for changes in things you care about. You say which companies, topics, or products you care about. The system spins up agents that continuously check relevant pages, feeds, or search results. When something meaningful changes, you get a short synthesized update. At some point, these agents may be able to take direct actions as well (buy item if price drops below certain amount, if competitor changes pricing, update our comparison page to reflect that).
Daily brief and inbox agents
There is a growing set of tools that connect to your email, calendar, and documents and then run every night or as new emails come in. Fyxer is a great example: it monitors your inbox, and autocategorizes email and drafts responses in your voice. Another good example are various daily briefing agents that run at scheduled times overnight and review the next day’s calendar and create
You can imagine similar patterns inside companies: agents that monitor various forms of data and proactively take action.
Closing thoughts
This cycle started by putting a chat box in front of a model. The next phase is about letting agents running on their own time - it allows them to run for longer, always remain running/monitoring and even be proactive.
Background agents are the mechanism that makes long running, multi step, and recurring work feel natural. They may start from chat, but they do not have to live there. They can be user-created, or ambient and always on. They can be quiet most of the time, then proactive in short bursts when something actually matters.
My guess is that most of the real leverage from AI will come from those quiet loops: agents working in the background while we are in meetings or asleep, and then appearing, briefly, with something that is actually useful. Chat will still be there, but more as the way we brief and tune them, rather than the place where all the work happens.
If you have thoughts about this paradigm or are building various forms of these agents, I’d love to hear from you at tanay at wing.vc