# Attention is all we have now How I manage my attention with tens of agents running at once, a checkpoint Canonical: https://agentrun.ai/blog/attention-is-all-we-have-now Published: 2026-11-05T19:49:14.000Z Author: Miguel Rios Berrios _How I manage my attention with tens of agents running at once, a checkpoint_ In a world where intelligence, compute, storage, pretty much everything around us is scaling rapidly, our human attention is going to be our scarcest resource, and the way we manage it will be one of the strongest indicators of success ![Sketch showing intelligence, compute and storage growing rapidly while human attention stays nearly flat.](/blog/imported-writing/attention-is-all-we-have-now/563e9a2ab45ac2e7.png) Last night I was at a dinner with other CTOs hosted by [Sherwood Callaway](https://x.com/shcallaway) from [Sazabi](https://www.linkedin.com/company/sazabi/) . One of the attendees asked how everyone was managing their attention span now that we're all running so many agents. The consensus was that keeping up with more than a handful of agents at once breaks your focus and doesn't feel sustainable. If you've followed me for a while, you know I've been on a quest to have agents run as long as possible while producing high-quality results. I felt early on that human attention wouldn't scale as fast as the intelligence we'd have, and that it would eventually become the problem. Now it's all we have, especially with agents doing so much of the work. About a year ago I [wrote about this](https://www.linkedin.com/posts/miguelriosberrios_%F0%9D%90%80%F0%9D%90%AD%F0%9D%90%AD%F0%9D%90%9E%F0%9D%90%A7%F0%9D%90%AD%F0%9D%90%A2%F0%9D%90%A8%F0%9D%90%A7-%F0%9D%90%A2%F0%9D%90%AC-%F0%9D%90%80%F0%9D%90%A5%F0%9D%90%A5-%F0%9D%90%96%F0%9D%90%9E-%F0%9D%90%87%F0%9D%90%9A%F0%9D%90%AF%F0%9D%90%9E-share-7394276979094618112-oTVK/), and some of it still holds. This is a checkpoint with updated advice. If you're terminally online and using Opus 5.5 every day, some of it won't be new, but I hope it helps if you're starting to feel what it's like to work with a lot of agents at once. ## What changed A year ago, Opus 4.6 marked the point where agents could do pretty much anything if you paid close attention. For me, Opus 5.5 marked the point where you can trust them to do pretty much anything with the right controls and verification. I've written about how we verify: [autoqa](/blog/from-qa-to-autoqa) for the product, hill climbing on metrics, and agents acting as guardrails for other agents. None of it is perfect, but it lets me run a lot of agents at once. This is also when I spend the least time in front of agent sessions, intentionally. I used to cycle through sessions in a TUI or a Mac app, working with each one. Now I'm involved in fewer of them and a lot more gets done. From July to October, I typed about the same number of messages each week. Agent turns went from about 7,000 a week to well over 100,000, and merged PRs went from about 160 a week to roughly 500, with one week over 900. I still work in about two panes an hour, same as in the summer. ![Weekly charts compare messages typed, agent turns and merged pull requests; agent output grows much faster than typing.](/blog/imported-writing/attention-is-all-we-have-now/62e9073803278e2b.png) The obvious question is whether the work got worse. As far as our code reviews can tell, it didn't. Since late September, PRs that agents took end to end with me mostly approving got the same first score from Greptile as PRs where people were in the loop, about 4.1 out of 5 for both, and about four in ten got a serious finding either way. The two dipped together at the end of September and climbed back together through October. ![Weekly Greptile charts compare reviewed pull requests, first review scores and the share with serious findings.](/blog/imported-writing/attention-is-all-we-have-now/ad70f9a17b240411.png) Here's how I work right now. ## 1\. More push, less pull Early on you had to do a lot of pulling: going from session to session, reading the diffs and tool calls. It felt like I had to be involved in everything. Now I'm less involved in each session and more involved at a higher level. The agents mostly run on their own and push to me through Slack or my phone when they need me: a large task is done, a goal is met, or an agent is blocked on my approval. Otherwise I don't need to be there. Another way to force push over pull is agents that run end to end on a schedule or a trigger. We have agents that read our logs and traces every day and figure out what our product's agent could have done better. If an agent is confident in a fix, it writes a PR that gets reviewed. If the PR is low risk, meaning it doesn't touch the money path, authentication or our data model, the agent takes it from triage to a release that a person deploys to production. I have fewer touch points, at the cost of knowing less about what's going on. We started with simple bugs, and with strong autoqa, CI, tests and code review it has grown to most of our changes. I'd say an agent can now discover, plan and complete about 90% of them. ## 2\. High-density time on what matters What matters is subjective, and I've calibrated it over a couple of years of building my company, so here's a typical day. I wake up, take care of my girls, make them breakfast and take them to school. Then I sit down at my laptop for my first block of high-density focus, usually about an hour while I wait for the next ferry. These days I write first (you may have noticed I've been writing a lot). Then I check in with a couple of agents, mostly the ones running large initiatives. One of them commands the other agents, and I ask it something specific: ``` I'm back. Give me a recap of the last eight hours. Where are we? What's blocked on me, what's blocked on something else, and what's next? ``` _The prompt I use in the morning_ Then I unblock what needs me: usually code to review, a decision the agent couldn't make, or something that needs more discovery or planning. After that another agent goes through our metrics and anything that happened overnight, like something failing or something that looks off and needs me to dig in. During this time I'm working with two or three agents at most. Plenty of others are running or waiting for a trigger, but I'm not involved with them, and I'm no longer cycling through five or more. The night before is what makes the morning work. Before bed, I tell the agent in charge what to get done overnight, to approve anything it's allowed to, and to leave anything that needs me for the morning: ``` I'm going to sleep. Keep unblocking as much as you can overnight and approve anything you're allowed to approve. Anything that needs me goes on one list for the morning. Don't stop to ask me anything: no AskUserQuestion, nothing that blocks you from overseeing the other agents. I'll let you know when I'm back. ``` _The prompt I leave at night_ Then I go to work for standup, which is usually where we help someone who's blocked or pick up new work from a customer meeting or a decision we just made. That leads to my second high-density block: planning anything new, usually 30 minutes or so with an agent doing discovery, exploring code and looking at possible solutions. For the rest of the day, I'm heavily involved in one or two things at most, usually a large change in the codebase, a new product feature or a customer. They come in two forms. Either it's how do I fix this fast, usually something critical that needs to be in production, or it's how do I get this off my plate, which means planning, iterating on a technical design, picking an approach and letting the agent cook. So I spend a couple of blocks a day on planning and verification, and the rest where I'm needed. We used to be involved in everything, from planning to development, testing and releasing. Now most of our high-density time goes to planning, decisions and verifying the work, and the agent does most of what's in between. Even the verification keeps getting better and more automated, with autoqa showing us product features before and after a change, evals to hill climb on, and skills for writing and code review that we can check. As models get better, I spend less time on how most things get built and more on how the few that matter get built, like the hardest problems in our product or anything that touches security. ## 3\. Work with people who operate like you This one has been extremely important, and I know it doesn't apply to everyone. At a large company it's probably harder, with more variety in seniority and work styles. I'm lucky to work with about six people who operate the way I do. We all use our agents like I described, so we talk less about things an agent can answer and more about what matters: open questions, design debates, and coordination across parts of the stack that need more than one person. We still talk all day, but the small interruptions (how was this done, where is this code, I'm stuck) go to our agents, which answer most of them with [recall](/blog/recall-how-a-skill-became-our-company-brain), our company brain. Our time together goes to problems that still need us. ## 4\. Get away from the screen while the agents work When agents run for a long time, you get to decide what else to do with yours. There's a gym in front of the office that I try to hit a couple of days a week, and I like going for walks. I'm "writing" this while walking the dog and talking to Granola. On weekends I go on at least a 50-mile bike ride and spend a ton of time with my kids. I've tried all sorts of crazy things to stay in the loop while I'm away, including having agents [call me](https://x.com/MiguelriosEN/status/2061186750924620023): I built a [hands-free skill](https://github.com/Parcha-ai/parcha-skills/tree/main/hands-free) with Vapi, so when an agent needs direction, it calls me. The work still gets done while I'm away. ![Heat map of hourly agent work above daily pull-request activity, showing increasing overnight activity.](/blog/imported-writing/attention-is-all-we-have-now/b9d988dcf3d3b9cf.png) What I'm trying to avoid is the overwhelm of hours in front of a screen, and today you can have a fairly healthy work style and still get a ton of high-quality work done. ## 5\. Keep removing yourself from the equation As models get better, and hopefully cheaper and faster, we'll feel the need to do more. There's a fair argument for slowing down and doing less, and for a lot of people it's probably the right call. The pressure to do more is still there, though: customers to support, competitors moving fast, or the anxiety of feeling like you have to keep moving. I'm not going to apologize for it, and I'm not signing up for doing less. I'm doing more and more, and I'm enjoying it a lot. My best advice is to find the bottlenecks that keep you babysitting your agents, and remove them. Get agents running for a long time with the work staying high quality, which takes very strong verification. Your bottlenecks may be different: if you do a lot of front-end work and need to point and click, figure out how much of that an agent can do. Then batch what still needs you into high-density blocks. For me, high density isn't multitasking. It's going deep on one set of topics, even through one agent. In the morning I plan across different lanes of work, and I balance that with stretches where one thing gets all my focus. ## A manager for your agents What has changed the most for me is having an agent that manages other agents. I've watched this emerge, and Opus 5.5 is the first model I feel does it well. In Claude Code it seems like post-training teaches the agent that it's managing other agents, and it handles that communication well even without the send-message tools, through a higher-level API like the one in herdr, which we use. I expect we'll see more of one session acting as the manager, or the captain the way [Capy](https://capy.ai/) does it. The downside is cost. It can burn through a weekly usage limit in a couple of hours, and at API rates it can run up to $1,000 a day. I'm not a believer in [token maxing](/blog/tokenmaxxing-won-t-survive-planning-season). Only do this if you're seeing the value or you need it. Otherwise, don't spend that kind of money on it. ## What I try not to do The main one is overseeing more than four or five sessions at a time. At that point you're not paying attention to any of them well enough to make a difference, and an agent would probably do it better. The exception is sessions so critical to your business that you shouldn't hand them to an agent. ## Where I want this to go Over time, I want my attention, which won't scale with more GPUs or more intelligence, reserved for the most critical things. Computers have a version of this. The L1 and L2 caches are the fastest memory, right next to the processor, and it would be great if they were huge. They can't be, because a bigger cache is a slower one and costs more chip area and power. So they stay small, and the hardware keeps them filled with what the processor is most likely to need next. I think the same will happen with our focus: instead of trying to scale it, we'll keep finding ways to protect it from more of the work. My goal is for the number of things that need me to go down while the amount of work we get done goes up. If you're doing ten things and only two need you, you got ten things done and only had to focus on two. I'd like the time I give to work to stay about the same or go down a bit, so more of it goes to riding my bike, my family and my friends ![Sketch of the intended trend: more work completed, fewer things requiring the author, and roughly steady human attention.](/blog/imported-writing/attention-is-all-we-have-now/4ffcc9f8b2f9b4f1.png) All of this will keep changing with the models, the harnesses and our tools. I'm looking forward to reading this a year from now and seeing how it held up.