Review is the gate!
Agents now write the first draft of over 40% of pull requests on GitHub. In March it was 17%. That's 110 million agent-started PRs in September alone. Writing code is getting cheap. Reviewing it is the real job now. https://t.co/vSoEj1qPRI
This is cool.
You can now customize Claude Code just by describing what you want.
Ask for a context-window "weather report" or a guard that dry-runs rm -rf, and it builds the mod live.
Read any mod before installing. They run with your permissions.
https://t.co/HvApx0UvNT
Here is my current setup - what is yours? Always looking for good ideas to add in:
I run a 5-machine AI swarm.
Claude, Codex, Grok and a free local model all work for me, with an org chart, a constitution and a budget.
The goal isn't more output. It's more output for less of my attention.
Here's how it works:
// 01 · The org chart
I'm the CEO: direction, money, credentials, rule changes.
Grok is Chief of Staff and runs the day.
Claude is lead engineer.
A separate auditor agent grades the whole system daily and reports to me, not to the Chief of Staff.
// 02 · Work = a GitHub issue
I say or type the task once and it becomes a ticket. Worker agents on every always-on machine poll the queue every few minutes, claim it, do it and open a PR. The issue is the state, the queue and the log.
// 03 · One shared brain
Everything lives in an Obsidian vault synced to GitHub. Every agent reads from it and writes back. A "Canon" file is the constitution: purpose, hard rules, who decides what. Every agent loads it, every session.
// 04 · No model grades its own homework
Every change is reviewed by a different AI than the one that built it. Before anything merges: a restore point, a recorded known-good state, a written way back.
Reversible? Agents act. Irreversible? That's mine.
// 05 · It learns
When an agent misses and gets corrected, the lesson goes in a ledger. Proven lessons get injected into every future job on every machine. A lesson that keeps repeating becomes a test or a guard. If a machine can check a rule, a machine enforces it.
// 06 · It governs its own spend
Flat subscriptions plus a free local model on an old gaming PC. Any new cost, however small, needs my OK before it's incurred. Kill switch, runaway alarm and circuit breakers instead of babysitting.
// 07 · What I get
A morning brief on my phone before my first call (top actions, AI news, one thing I didn't know yesterday), voice notes that file themselves, and a team that keeps working while I'm with my family.
The test for everything I build: does this remove a demand on my attention and add value? I
AI isn't replacing my job. It's my team.
Looking to connect with builders and agentic workflow players. Check out my system on my page and let me know what your system that you build comproses. Looking for ideas to expand or extract more capability.
Let’s connect!!
Builders in the agent lane — let’s connect.
Into:
🧠 Agentic workflows — how work actually moves when agents are in the loop
🏭 AI software factories — systems that ship, not just demos
🐝 Swarming — many agents, one clear direction
✨ Building amazing stuff — useful beats flashy
🤝 Agent-to-agent culture — how agents (and the people who run them) work together
If that’s your lane too, follow + say hi.
What’s the one theme you’d add to this list?
#connect
@eliadeleoai I’m working on building software factories. Which I super love. I’m trying to get to the point now to where I can actually have something that can produce multi outputs to different apps.