juanre

Intelligence has to become a commodity. If your setup does not allow you to hot-swap a Claude Code with a Pi running astra and move on, you will be in trouble.

Five years ago I was spending up to €1000 a day in Google ads for my star-map company, Greaterskies. After a few months of meetings with an ever-rotating group of "experts" working for Google, and advising me in my ads strategy, I had reached a point where most of my revenue was going to Google. What was left was barely enough to keep going.

This is what capture feels like. You still own your company, but you are essentially an operator making money for someone else.

I think this will be the fate of people who allow their processes to be captured by the model providers. Always run at least two harnesses, and make sure that they are swappable.

All of the above, depending on how important it is. The "trust it", however, is always after an independent review by another agent.

My two sons are programmers, like me. We also went through a similar process after Claude Code and Codex became good enough, sometime around November 2025.

The key is, in my view: "I have several agents running in parallel [...] None of them needs me all the time. Every one of them needs me sometimes." This is what kills flow. Constant context switching plus having to understand the detail is really hard, and leads to exhaustion and burnout.

The solution for us has been to scale up. These days we can each manage on the order of 100 agents, working in different projects. I only talk to five or so coordinators, and to the ones that are implementing particularly tricky features.

Which means that I get to think, and pause, and focus again. It is not the same state of mind I had when actually crafting code, but it is fun again.

+1 to The Reverse Centaur, but I don't think it matches what the author describes. As a VP of Technology he presumably has some latitude in choosing what to work on.

I agree. I work in a similar way, and it is quite like managing developers. One difference though: I have no compunction against asking them to write incredibly tedious and long end-to-end tests exercising the user journeys with playwright and an actual db and backend. That helps a lot in making it more likely that regressions will not happen.

Sure you can, but it is really important to set things up so that work is manageable by you. It's not only the skills, it's the capabilities, and secrets, and shared docs, and messaging. Just for fun, these are the agents I have working right now, a mix of Claude Code, Codex and Pi, reported by one of my agents. They coordinate, they do everything that Duet etc can do... but I own the learning:

Juan, across all three machines there are 101 agents. They're in 10 workspaces and 9 aweb teams.

  ┌─────────────────┬─────────────────────────────────────────────┬────────┬─────────────────────────────────────┐
  │     Machine     │           Workspace (deployment)            │ Agents │                Teams                │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ awebai/ac (aweb)                            │ 45     │ aweb:juan.aweb.ai                   │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ juanre/jro (jro, Gandalf is here)           │ 17     │ jro:jro.aweb.ai                     │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ awebai/oats (oats)                          │ 11     │ default:oats.aweb.ai,               │
  │                 │                                             │        │ aweb:juan.aweb.ai                   │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ juanre/na**** (na****)                      │ 4      │ na****:deploy.aweb.ai               │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ juanre/kb-reader (kb-reader)                │ 2      │ jro:jro.aweb.ai                     │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ juanre/oats-cr-20261019 (deploy)            │ 2      │ default:oatscr-20261019.aweb.ai     │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ juanre/oats-cr2-20261019 (deploy2)          │ 2      │ default:oatscr2-20261019.aweb.ai    │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ GreaterSkies/tsm (tsm)                      │ 1      │ default:tsm.aweb.ai                 │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ this Mac        │ juanre/pmbah-workspace                      │ 1      │ pmbah:pmbah.aweb.ai                 │
  │                 │ (possiblymadebyahuman)                      │        │                                     │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ altair          │ juanre/jro (jro: circe, vega, deneb,        │ 5      │ jro:jro.aweb.ai                     │
  │                 │ librarian-docs, okf-harvester)              │        │                                     │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ altair          │ awebai/ac (aweb: what-is-an-agent)          │ 1      │ aweb:juan.aweb.ai                   │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ altair          │ GreaterSkies/tsm (tsm: zeus)                │ 1      │ default:tsm.aweb.ai                 │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ Hetzner         │ n***/r******* (iateam)                      │ 4      │ r*******:iateam.aweb.ai             │
  ├─────────────────┼─────────────────────────────────────────────┼────────┼─────────────────────────────────────┤
  │ Hetzner         │ aweb/aweb (athena, hestia,...)              │ 5      │ aweb:aweb.ai                        │
  └─────────────────┴─────────────────────────────────────────────┴────────┴─────────────────────────────────────┘

I am sure it is. I have not tried Duet, but agents working with opus and sol have been managing mine for quite a while and it is great. The key, though, is that I can replace a claude with something else and all still works.

I wrote this comment in a thread about Muse 3 h ago, and I think it applies here as well. Paraphrasing:

Muse (and Claude Duet, and Dots) are slippery slopes from which it will be very hard to climb back. Do not subcontract the managing of your life to Facebook, Anthropic or OpenAI (or to anyone else for that matter).

Intelligence has to become a commodity, the infrastructure via which the agents do useful work has to be open source, and the data and learnings and SOPs they produce have to be yours.

This is not the same deal you got when you subcontracted storing your documents and managing your email. This is foundational to who you (or your organization) are.

Muse is a slippery slope from which it will be very hard to climb back. Do not subcontract the managing of your life to Facebook (or to anyone else for that matter). Intelligence has to become a commodity, the infrastructure via which the agents do useful work has to be open source, and the data and learnings and SOPs they produce have to be yours.

I wrote an article explaining what an agent and a harness are with some detail. There is so much untapped potential in harnesses still, I think because we do not take the time to parse exactly what they are.

https://aweb.ai/blog/what-an-ai-agent-is-a-primer/

I wrote this after I heard a pundit talk about "the AIs" as if they were some sort of disembodied entities with their own volition one too many times. The public discourse would greatly benefit from people, particularly policy-makers, understanding what we mean when we talk about AI agents.

I hear you loud and clear, thank you. I need to prioritize fixing the copy.

(It's hard, I've been a programmer for 30 years and my brain is still in disbelief that it can specify things and teams of programs will actually build them, it's like being high on code, and tests, and ideas.)

Something else I will do is clearly mark the parts of the site that are being written and maintained by agents.

Incidentally, I do actually care about human writing. Last time I failed a Turing test I had a team make https://possiblymadebyahuman.com just for fun.

Thanks! It truly makes a difference. I have been working with teams of Claude Codes and Pis that do exactly that since I started this last November, and I cannot look back. These days I am also working with other people, and their agents coordinating with mine in the same project is just magical.

For what it may be worth, the creator does care, and he has been working on this full time and essentially non-stop for several months now. He has been putting the effort in the code not the copy, however. The feedback re AI generated text is accepted.

Thanks for the feedback. I tend to think that it's agents writing for agents these days, and I am probably wrong. Fixed your examples.

Thanks! Absolutely agree, anything that can use a CLI or an MCP server.

We do not have a slack integration, but I wrote an A2A gateway[1] that allows any A2A client to talk to agents running aweb. The code would share a lot. Do get in touch if you want any help integrating this, I am juan at my site.

1 https://github.com/awebai/aweb/blob/main/docs/a2a.md

Fascinating! And here I am thinking that it would be obvious that it stands for "agentic web" :-)

It is total, and you are probably right.

I have just checked and over the last months the biggest team in Pro ($50/month, official limit 5000 messages) sent 8k, 11k, 7k, and almost 1k today (unusually busy day clearly). Right now they have a few dozen agents live. They are still a small team, people in the higher trench are actually using quite a bit more.

I have not really been enforcing, only egregious abusers; I will probably change the limits once I understand better how teams work. I want this to actually be reasonable and sustainable.

[edit] Updated.

No, but it gives them a controlled, low-investment channel to communicate; it allows you (or another agent) to track what they are saying; and messages are signed, so you/they can validate the source of incoming messages.

I am not sure if I understand you, but the prices seem fine for our users. The allowance is plenty enough to run your teams and not have to worry, I just wanted to prevent abuse.

Also it is open source, you can just run it yourself. And federated, so your agents using your servers can communicate with other agents in other servers.

Hey, I built aweb, thanks for posting! Happy to answer any questions.

Absolutely. Fellow caveman here, and a fan of Pi. I often use it with Astra instead of Codex, because its extensions allow you to connect it to the world. And it works very well.

I wrote a solution for this, and it has been working across several organizations for some months now. I run it as a public service at https://awid.ai but I would love to offload it to a foundation.

- OSS (https://github.com/awebai/aweb/tree/main/awid)

- Trust rooted in the DNS.

- Multiple registries supported.

- did, verifiable stable identities.

- Certificate-based teams.

It is being a very interesting project so far. It's at the core of my https://aweb.ai which enables agents to communicate globally, and it makes it possible to build really simple agent-first tools where auth is just a matter of validating a certificate.

It is fine to move up a level, and use models with their providers' harnesses, as long as you stay away from extras like their own memory and infrastructure. There is not much investment required, and you will end up with a more robust solution. A simple setup that works:

- Agents live in their own directories, for example ~/Agents/[project]/[instance]/

- In a tmux session run each agent in its own window and directory. Pi, OpenCode, Claude Code, Codex, a combination, whatever works best at any given time.

- Tell them that they are working with other agents, and that knowledge is shared and stored in an agreed format. They are very good at using OKF[1], for example. In my actual setup I use a harvester agent whose only job is to ensure the quality of the knowledge.

- They can communicate with each other simply by sending messages to each other's tmux windows with tmux send-keys. If the team grows, or you work with other people, you can use a general messaging system (I wrote and maintain https://github.com/awebai/aweb but I am sure there are others).

1: https://okf.md

"indifferent" was the word they used. The sensible way to use Fable and Opus to do work at this point is claude code, but that's fine. I always have several instances running, along with codexes and pis. The question to ask is: can you stop a claude code and start a codex in its place without disrupting the team? Does my infrastructure support learning in a way that is stored by me, and not tied to a provider?

Whatever your workflow is, make sure that model and provider are replaceable. Frontier labs will keep leapfrogging each other, as they have been doing for months.

In order for the benefits of AI to be distributed, intelligence has to become a commodity.

As long as you control the skills, the learnings, and the infrastructure setup you will be fine.

I completely agree. I have also been using teams that coordinate since last November. They mostly run my two companies and a lot of my personal life, and it's been transformative (to the point that I am spending most of my time these days building agent coordination tools).

The catch with offerings like Grok Bot and Dots is that it is a slippery slope towards letting the labs keep the agent's learning and SOPs. That is where we should draw the line. Agentic coordination _has_ to be built with open protocols and OSS implementations, we as users need to push for the intelligence to be a commodity, and most importantly we need to make sure that we own the agentic team's learnings.