My fleet now has 3 points of entry/stops for me to weigh in:
1. Objectives - I set the North star and explicit objectives.
2. I review (proposed worker/fleet) plans but I'm heavily biased towards 2 buttons here - approve/decline and send back to re-review against objectives.
3. Approve and deploy PR - although I think this can be remove too.
So essentially 1&2 are kind of needed now, but I reckon soon numbers 2 & 3 will be automated too!
2
149
Building an automated harness is quite interesting but also mentally taxing. It's like trying to keep fifty things in the head at the same time and connecting them. I reckon it's all about an iterative approach to it. ship harness v1 and then probably it will burn a ton of toks, then keep iterating while maximising based on the metrics and goals and minimising token spend. What a rollercoaster!
1
76
👏🏻🤙
All the people on here always tell me for years "nooooo you can't just ship solo and have high uptime like big companies"
And it's just not true, because big companies like Revolut even make stupid mistakes like not having monitoring working on their main production databases! And big companies go down all the time as you saw with Microsoft this week too
The point being is that anyone can go down, and it's pretty equal for a well managed solo founder as it is for a big company
1
64
I am seeing emerging trends where people build their own “family” agents where these are used to text and communicate with relatives and other not very close friends to keep their relationships warm.
A bit hardcore imo. I am pretty certain the people on the other side don’t know this is a bot.
1
2
28
Grok is amazing at voice and orchestration and it runs behind a computer so can nicely ssh into my hetzner box. ✅
I wondered if Claude can do that - from my phone. It couldn’t since it can’t ssh.
So what I did is make my own custom mcp server and run it on that same hetzner box via caddy (super light) then secure that url (via a bearer token).
Then go to claude on my phone and add that custom mcp and voila i can now talk to my hetzner claudes running via tmux on my hetzner box with voice from my phone.
More PRs!
3
6
150
I now see that one should have skills as versioned artefacts. All data store activities should then link into the version of a certain skill. This way an eval process can be set to have skills scored and improved based on actual data recursively.
Week 1 test skill a v1 and record all actions with this tagged version id.
Week 2 measure effectiveness of said skill against real data for week 1 then try an experiment or a change and make skill a v2. Iterate and experiment more.
Set claude’s objective to maximise a certain metric in that specific skill!
Combine skills and put local and global objectives!
🍺
2
48
At this pace its worth just nuking .MD files and orchestration setups once a new model arrives. And let it rip through the setup and offer a better one. It's like a version upgrade to the internal "OS" that runs the fleet of agents. Doing it now - curious how much better this works, but I noticed it is more of a mental shift move rather than operational move.
1
29




