Always-on agents - agents that work all the time
Andy Avramenko 16. novembar 2026 4 min čitanja 0 pregleda
Anthropic published an article on building always-on agents. All good, except the definition. A lot of what I have is built around one Claude Code session.
Anthropic published an article on how to build always-on agents. Everything in it is good, except the definition. If you take the idea that an always-on agent is started by a webhook, a background task or a schedule, then it's not very clear why just opening Claude Code and typing a prompt is not always-on. In principle, any agent you can start is always-on :) Well, except an agent that no longer starts, that one becomes always-off :)
A lot of what I have is built around one single Claude Code session, which is a bit closer to always-on as I understand it, but it comes with a lot of nuances I have to deal with:
1. The first, obvious one is the context window. We recently learned to solve it with a hook. When it reaches some high value (300k?), Claude starts picking a good moment to restart itself. It saves everything important and types /clear to itself.
2. I've also written before that, on principle, I want any more or less serious task to always run in the project folder (without the shared CLAUDE.md that lists all the projects, how to talk to me and so on), on a machine that has enough weekly limits. We've learned to solve this perfectly.
3. A separate art is the "next best action" - while I'm away, everyone should be busy. Some things are easy to handle on a schedule or on an event, but sometimes the nuances matter. For example, I want to build an AML bot, but across my four Anthropic subscriptions the weekly limits forecast is negative. If I move this task to next week, nothing bad will happen. Meanwhile we can run some pilot or have a call with the industry heavyweights. There are many of these constraints - there might even be some outage and everyone is fixing it...
This is solved with a pulse of initiatives. Every hour from lunch until midnight a simple script goes through the list of initiatives and asks the session - "what do we have here? what can we do?" Insanely useful, because you often launch something, especially something small, and forget about it - and it's awkward to use, or the price is bad, or it should be talked about on LinkedIn.
4. And then the morning rituals with me. Usually by voice. I still don't want to do absolutely nothing, so for the first half of the day it has to keep me entertained - tell me the news, what good happened in the business, what the priorities are, some decisions/inputs, plan talks and conferences, German, sport in the breaks... Every time between a couple of tasks there is an automatic context clear with a handoff between sessions.
In the end, an agent built like this still has no real natural agency, it doesn't sit there on its own going "damn, what useful thing could I do"... But with a well-built schedule and the company of other agents it is more or less busy with something useful. And this is even closer to the "always-on" use case itself, because it has to coordinate the work of other agents and processes, it just steps out for some fresh air and comes back with a clear head :)
Everything I have is built for myself, one day I'll make a demo. Slava Sazhin managed to make something universal, and it seems to be getting closer and closer to the benchmark. His agents have less poetic names, but judging by how much tennis he plays and how much he travels, it works even a bit better :)
Da li je ovaj članak bio vredan vašeg vremena?
Pročitajte i ovo
Cortisol
What AI agents can borrow from living nature besides ants. By default, an agent needs nothing. The job of a smart agent is to synchronize rhythms.
When the job is too big for your Claude Code server
Some builds need more memory than your machine has. Now Claude Code can borrow a bigger machine for the job: 3 cores and 6 GB, deleted when it is done.
More on ways to interact with AI. How I am learning German
Voice is not the only alternative to a monitor and keyboard. Claude makes a PDF of the morning mistakes and sends it to the printer. I fill it in by pen.
Sopstveni računar za Claude
Vaš sopstveni Linux računar sa instaliranim Claude Code, koji radi danonoćno dok je vaš laptop zatvoren. Od $3.49 mesečno.