Building something new. Self-taught designer and dev. Product guy, formerly of @revolut.
Joined January 2009
- Tweets27.6K
- Following1.9K
- Followers2.6K
- Likes21.3K
This is cool! Long-term maybe it’ll be a standard protocol like ACP/MCP.
Persistent agent networks are probably the future, but it feels so early & rough right now.
Today, you can jerry rig an agent network over Slack. But creating a Slack app has a ton of footguns: permission scopes, auth tokens, how to wake an agent, and how agents can prove their identity.
(e.g. Grok Bot can wake over webhooks, and Dots’ can wake on @ mentions, but Claude has to poll Slack on a routine)
Some models (like all of OpenAI’s ones) are also bad at meta-prompting, so they can’t be an equal peer in the network and has to be a subordinate to another.
Seems to be the way it’s all trending though. I bet paseo/herdr/conductor style orchestration will seem quaint in a few months.
Let your Grok Bot / Muse talk to your Claude Code / Codex or any other agents
Introducing Reason Network,
Free, peer to peer and Open-Source Engine
Loop in your teammates too so no live agents work collides when coding.
Try it for free: reasonmachines.com/network
one click install and setup!
Made a Grok @Bot for watching @thsottiaux resets, so you can drive your software factory in reaction to whenever a Codex reset is impending or has already happened.
To use this, hook up your Grok Bot to a script that watches for your own Codex usage, and a tool that drives Codex sessions headlessly (you can use the codex CLI, but I prefer Paseo).
Then you can just give Grok Bot work to burn each reset down with.
x.ai/bot/e5T5aotYdq4p1etGEVv…
One major plus that Grok Bot has over Claude Projects and Dot is how agentic it is.
Both Claude Projects and Dot really loves telling you to do stuff for them, as if you are their agent, instead of the other way around.
They are often blocked by some hard safety wall that cannot be bypassed.
Multiple times a day I am told to manually approve permissions, manually type commands in the terminal, manually click around the browser, and so on.
This is a serious regression in Claude and Codex, as much as their models are powerful, it’s become increasingly frustrating to use them for any sort of hands-off autonomous work.
Meanwhile, Grok Bot just does things, as you expect an agent would.
It’s not doing it with wild abandon either.
It’s still cautionary, but respects when I tell it to do something.
I’m now on the $100 ChatGPT plan and the $200 Claude plan, and neither product works agentically as promised.
An “always-on agent” that mostly just tells me to act on its behalf multiple times a day (sometimes multiple times an hour), litters my repo with a sprawl of worktrees despite instructions to the contrary, and refuses to clean up the mess it makes because of “safety checks” which sometimes can’t be bypassed with explicit instruction, is extremely frustrating to work with.
When the bar is lowered in one way, it also gets raised in another way.
Such that it all nets out.
Usually the complements of it gets more difficult or more expensive, especially the competitive parts of it.
So it’s easier to make apps now.
But it’s also harder to buy a suitable domain name, harder to find a name not already trademarked, harder to secure your app, and harder to compete for attention.
Like with digital photography, desktop publishing, and electronic music, the more people can make something, the more of it is made for casual and personal reasons
But the harder it is to make it for commercial reasons.
Saying software is dead because anyone can build it is like saying photography is dead now that everyone in the world has a camera in their pocket
Not only is it not dead it's the beginning of the Kodak Brownie moment for software, and we will see an absolute explosion of it
It will be the first generation for who building software will feel no different than using it. This is crazy!
The quality and bar of what we expect from our everyday apps will be raised dramatically, we'll have personal software, we'll have disposable software, and agents will be the foundation for most of it
Tech has always been chaotic and evolving, the people who have been able to adapt and move in front of new opportunities have always excelled, this is nothing new, I'm seeing it now and we'll see it 10 years from now
Find the opportunities in the chaose and disruption, there are countless of them
The folks at OpenAI need a lesson in economic incentives.
They taught their entire userbase to burn usage as quickly as possible, because it could be reset at anytime without warning, leaving money on the table.
This incentivized spiky usage, which the existence of usage limits were meant to prevent in the first place.
I’m not sure if they really meant to incentivize spiky usage, but, if they didn’t, they’re in a rough spot.
It’s not a good look when your head of product is publicly signaling confusion about a pricing policy that should have been intentional.
Do you want to incentivize spiky usage (global resets) or do you want to discourage it (5 hour limits)? Make up your mind!
I would bet a lot of money that this is directionally correct but specifically wrong
in that I think, yes, the archetypal “indie hacker” is dead in the water
Why pay an indie hacker when you can have ~infinite code monkeys in your pocket for $20–200?
But I think this post misses an emerging archetype that will likely succeed as a solopreneur, supercharged by coding agents.
And that’s the jack-of-all-trades product engineer, design engineer, or PM that can now ship as much as a small team at a startup.
Might not seem like there are many such individuals right now (since there’s barely any upside to “building in public” anymore)
But I’d bet that by next year or the next, we’ll see many ex-unicorn or ex-bigtech folks start launching spectacular products – without a team and without VC.
There are legions of such talented people from the 2010s tech boom who are post-economic, FIRE’d, or laid-off-and-coasting.
Opus 5.5 only just made it possible for these people to go for it.
I’m less bullish than I used to be about the future of solopreneurs. In fact if you look closely some of them are already pivoting to become more like creators.
On the one hand they can now build almost anything pretty fast. On the other hand simple products are the ones most at risk of being disrupted because there's no moat. The really complex ones that can't be copied easily need a team and resources to be built.
There will be a few outliers that will make a lot of money. But I find it hard to imagine it being a new economy that replaces the traditional job for the average guy.
Gotta say, the coordinator pattern of working with coding agents is growing on me.
Feels more “agent-native” than working with pinned threads in a sidebar, or with an orchestrator like Paseo.
The UX is still rough though:
Cascading permissions, open decisions, and blockers are still a huge pain. I still have to go into the child threads to unblock them, more often than I’d like.
The coordinator agent is often arbitrarily hobbled by lack of capabilities, compared to bare threads started in Claude Code, Codex, or Cursor:
- Claude’s project coordinator has no web browser and can’t use MCPs
- Dot can’t run skills or see your AGENTS.md or use the structured user question tool. It also can’t use your local computer or browser directly
- Grok Bot doesn’t use your repo’s skills or AGENTS.md
The threads launched by the coordinator are also themselves arbitrarily hobbled by lack of capabilities!
For instance, in Claude Projects, you can’t invoke skills by hand inside project threads, only inside the main coordinator. You also can’t see what skills that thread has read.
Been trying out Claude Projects, ChatGPT Dots, and Grok Bot as a coordinator for coding agents.
A missed opportunity of all of them is how they can’t use the repo’s existing AGENTS.md, MCP setup, or skills.
A coordinator agent should not have less context than the threads they’re coordinating.
Tried Slack again after years of not using it. Can’t believe how bad it’s gotten.
Everything is a mess, and you can’t tell which threads you’ve read or not read.
There are three or four ways to check your unread messages, and they all show a different set of messages and bombard you with notification badges.
How do big companies even survive communicating with this infernal tool?
Where are the companies vibe coding themselves Slack alternatives? I expect a lot of this in 2027
The big question imo is the replacement for Slack Connect. I wonder if Cambria by @geoffreylitt – bidirectional lenses – is the right idea for that
This is a welcome improvement!
Coding agents used to eat away at your SSD and battery life by writing TBs of SQLite logs and GBs of session rollout files as they work.
Glad this is soon to be a thing of the past!
Grok Bot’s worked this way since day 1 and it’s always felt right to me: the harness, the thread/session, and the sandbox are in the cloud.
You can update the app or restart your computer without work suddenly stopping.
There’s no compromise because the agent can still work on your computer locally and run shell commands or manipulate files.
Best of both worlds!
ChatGPT Dots vs Grok Bot Impressions
- Dot needs more handholding and doesn’t “get it” as much as Grok Bot. Surprising, given it’s powered by Astra.
- Dot has an overkill safety policy, some of which you can’t even override. For example I have a workflow which sends my code to Typesafe’s API for Jev to lint my code against some heuristics, and it flags it for my manual permissions every time.
- Dot can’t interact with the rest of ChatGPT via tool calls at all. It can spawn and manage Codex threads only. Grok Bots on the other hand can do almost everything on your behalf that you can do in the UI.
- Dot drops the ball about as often as Grok Bot. But is more of a pain to see what it failed to follow up on, because you can only have one Dot, but you can have multiple Grok Bots and groups. The workaround is to install it in Slack and talk to it via different channels (which works, but it should just be built into the app).
- Unlike Grok Bot, there’s no transparency as to what routines, skills, or memories Dots have. Makes it hard to audit where things have gone wrong.
- There’s an easier Slack integration than Grok Bot (one-click setup for adding your dot as an agent to Slack)
- You can’t setup custom webhook-based routines in Dot like you can with Grok Bot. MCPs can trigger events, but it depends on that MCP to support it.
- Grok Bot can directly interact with your local computer and run scripts like a coding agent. Dot can do that only via a Codex thread that it spawns (which clutters up your Codex sidebar and eats into your Codex allowance)
- The remote desktop in Dots is faster (feels almost instant like Windows RDP), whereas the one in Grok Bot is slower and feels like using VNC.
- Dots is more proactive out of the box than Grok Bot (it looks at your connected apps like your email and asks you about possible things it can do), but then it isn’t actually proactive once you start using it. You can tell it to do something (like send you a routine update) and it will acknowledge, then happily forget about it.
- You can’t mark messages as unread in Dots, whereas you can in Grok Bot. Combined with having multiple bots, it makes following up on different things easier on Grok Bot. Again, you can workaround on Slack, but that’s another service to pay for if you’re not already using it.
- Both Astra and Grok 4.7 speak in a way that isn’t human-friendly by default, and while it isn’t as grating as speaking to Opus 5, it can get on your nerves after a while. I’m not sure why Astra speaks in this way in Dot: in ChatGPT (6 Pro), Astra feels pleasantly conversational (perhaps it’s my custom instructions).
- Speaking of ChatGPT custom instructions, it’s unclear whether those custom instructions is used by Dot at all.
Rethinking my subscriptions after DevDay
Feels like this is the best path forward, balancing cost & frontier performance:
- Claude Max 5x ($100): buys you access to Fable 5.1 (the best architect and code reviewer) and Opus 5.5 (the best daily driver)
- ChatGPT Pro 100 ($100): buys you access to GPT-6 Astra Pro (the best planner/deep researcher), Codex Security Review, and Luna/Sol as spare capacity when you run into Claude’s 5h limits, and Dots as a smart orchestrator agent
- Cursor Pro ($20): buys you access to Grok Bot (the best UI for automating routine work), and bonus usage of Grok 4.7 which is useful for offloading dumb routine and maintenance tasks.
Works out to be $220.
If needed, bump Claude to Max 20x ($200) for more daily driver use, or Cursor to Pro+ ($60) for more routine automation.
So $300–$360 at the most.
Still clocks in at less than $500 for the new OpenAI plan!