This week, Anthropic shipped Channels, letting developers message Claude Code from Telegram and Discord (docs). A few days before that, Dispatch, which lets you control Cowork from your phone. Before that, Remote Control. Three features in the space of a month, all pushing the same direction: AI that runs in the background, waiting for you, while you do something else.
Every AI company is racing to get here. OpenClaw proved the demand. The pitch is simple: your agent works while you walk the dog. You come back to finished code. The future of development is asynchronous, persistent, always on.
And it is. I've been living some version of this for months. I've got agents and automations handling chunks of my development workflow, my research, my admin. It's not a future state for me. It's a Tuesday.
But nobody's being honest about what it actually feels like.
The Vampire Problem
Steve Yegge wrote a piece in February called "The AI Vampire" that's been rattling around in my head ever since. His argument: AI genuinely makes you more productive, but that productivity gets extracted. By your employer, by your own ambition, by the tools themselves. You do more because you can do more, and then you do more again, and at some point you look up and realise you're exhausted but you can't explain why. You were "just prompting."
Around the same time, UC Berkeley researchers published a study in the Harvard Business Review. They spent eight months inside a 200-person tech company and found something uncomfortable: nobody was being forced to overwork. People voluntarily expanded their workload because AI made "doing more" feel possible. The burnout wasn't imposed. It was self-inflicted. The tools made more feel achievable, so more became the baseline.
I recognise both of these patterns. Not because I've burned out, but because I've had to actively, deliberately build my way around burning out.
What Nobody Tells You About Orchestrating Agents
Here's the thing about running multiple AI workflows in parallel: it doesn't feel like coding. It doesn't feel like managing, either. It feels like something new that doesn't have a name yet.
There's a rhythm to it. You set an agent working on something, turn to another task, check back, course-correct, set it going again. At any given moment, there's usually one agent that needs your input. The others are either working or waiting. You're never entirely sure which without checking.
And that's where the new kind of tiredness comes from. It's not the tiredness of writing code for ten hours. It's the tiredness of holding multiple threads in your head, context-switching between them, and making a constant stream of small decisions. Every check-in is a micro-judgement: is this output right? Is it heading in the right direction? Does it need correction or can I let it run?
It's cognitively expensive in a way that's hard to see from the outside. You're not typing much. You're not in meetings. You might be sitting quietly with a coffee. But your brain is doing something closer to air traffic control than programming.
And then there's the guilt. The strange, specific guilt of having nothing running. If the agents are idle, you feel like you should be feeding them. If you're not prompting, you're not being productive. The tool has created its own sense of urgency, even when there's no deadline.
This is the vampire, and it doesn't need your employer to work. It works fine on freelancers sitting alone at their desks.
How I've Made It Work (So Far)
I'm not writing this as a cautionary tale. I'm writing it because I think this is genuinely the most productive and interesting way I've ever worked, and I want to be honest about the bits that are hard.
What's made the difference for me is that I've approached it methodically. I haven't just thrown AI at everything and hoped for the best. I've taken each part of my workflow, built specific agents and automations for it, and applied my own guardrails, preferences, and quality checks. Each one reflects how I actually want to work, not just how the tool defaults to working.
That last bit matters. The defaults will drain you. If you let AI set the pace, the pace is infinite. There is always another prompt, another task you could parallelise, another agent you could spin up. The tool will never tell you to stop.
So you have to build the stopping points yourself. For me, that means structured workflows with clear checkpoints. It means knowing which tasks benefit from AI speed and which ones need me to slow down and think. It means accepting that sometimes the most productive thing I can do is close the laptop and go for a walk, even when three agents are sitting idle.
Yegge suggests the new workday should be three to four hours of this kind of high-intensity, decision-heavy work. I think he's probably right, at least for the orchestration part. The rest of the day can be human work: conversations, thinking, planning, the stuff that doesn't happen inside a terminal.
The Bit That Matters
The conversation around AI and work keeps getting stuck in two camps. Camp one: AI is making everything better, we're all 10x now, the future is here. Camp two: AI is destroying jobs, burning people out, the sky is falling.
Both are missing the point, in my opinion. The interesting thing isn't whether AI makes you more productive. It does. The interesting thing is that it fundamentally changes the texture of the work, and almost nobody is talking about how to navigate that well.
The features shipping this week, Dispatch, Channels, always-on agents, they're going to accelerate this. More developers are about to discover what it's like to have AI running in the background while they live their life. Some will thrive. Some will burn out before they understand what's happening.
The difference, I think, will come down to how deliberately you approach it. Whether you design your workflow or let the tools design it for you. Whether you bring your own judgement and experience to the process, or let the pace of the tool become the pace of your life.
AI is a tool. An extraordinary, genuinely transformative tool. But a tool that can't drive itself, and one that will happily run you into the ground if you let it.
Build the guardrails before you need them. Your future self will thank you.