Suat Özkaya | AI & Mobile retweeted
Your agents worked all afternoon. What did each one ask to do, and who answered it?
ACTower's decision log keeps every request with its own ID, per pane, and who answered it: ACTower's rules, Autopilot, you in ACTower, or you in the terminal.
It stays on your machine.
Another must watch from @mattpocockuk
It's been nearly 6 months since I posted a 'watch me work' deep dive on my skills.
Since then, they've become the most-starred skills repo in the world.
So, here's 70 minutes of me building a 21,000-line PR, especially focusing on:
- DDD with /grill-with-docs
- Improving your codebase with /retro
- How to do human review effectively
Enjoy!
21
ACTower is getting better with each release 🎉
Restarting a long Claude Code session by hand: ask for a handoff, start fresh, paste it back.
ACTower does all three: past your token limit, after ten quiet minutes, it asks for the handoff, restarts Claude Code and gives it back.
Off by default. Each auto restart is logged.
2
33
Suat Özkaya | AI & Mobile retweeted
ACTower 1.3.7 is out.
The Responder, redesigned around what needs you: your questions on top, Autopilot's folded below, busy panes one click away.
Also: a calmer Auto Restart, and Autopilot answers more on its own.
Suat Özkaya | AI & Mobile retweeted
Fable 5.2 won't take your job.
Fable 5.2 will take everything from you.
Today I was working with Gemini 3.8 Flash in @antigravity, with a TDD skill loaded.
The model read the skill, then quietly skipped it.
@googleaidevs a short thread on what happened 🧵
2
61
Model or harness? My read: mostly the model.
The skill loaded fine. The model knew the rule and still broke it.
But the harness isn’t fully off the hook: a skill is just text in context. Nothing enforces “no code before a red test.”
1
4
Takeaways:
Loading a skill ≠ the model following it
Verify the process, not just the output
Ask for evidence: failing test output, order of changes, etc.
One experience, not a benchmark. Curious how the @antigravity team thinks about enforcing skills in the harness.
14
Suat Özkaya | AI & Mobile retweeted
Some agents in tmux, some in #herdr. Now you have two places to check who is waiting on you.
ACTower works with both: their questions land in one queue, safe ones get approved automatically, the rest wait for you.
Keep the setup you like. Where do your agents run?
Suat Özkaya | AI & Mobile retweeted
Agents shouldn't start every session from scratch.
ACTower Ecosystems give agents and humans a shared organizational memory across repos and sessions.
Agents read it before each task and add to it after. Humans read the same pages in Obsidian.
actower.io/docs/ecosystem
Suat Özkaya | AI & Mobile retweeted
ACTower 1.3.4 is out.
The control tower for your AI coding agents: it approves the safe questions for you, runs whole tasks on Autopilot with its own reviewers, and lets a team work from one shared ecosystem with proposals and reviews.
What's new below.
True
Agents need narrow permissions, protocols that work together, and costs measured per completed task.
At AGNTCon + MCPDev Summit EU, maintainers from @AnthropicAI, @soloio_inc, @Google, and @OpenAI joined @Gartner_inc's Nitish Tyagi to discuss what it takes to build and operate agentic systems.
The discussion covered tool loading, tracing why an agent acted, and testing standards in working implementations before changing the spec.
Read the recap: aaif.io/blog/maintainer-roun…
20
I’ll try this one, had a similar idea but no time to work on it.
I built Petal, a free, open-source disk-space explorer for macOS.
The chart draws itself while it scans. It flags what's safe to delete (Xcode builds, node_modules, caches) and tells you exactly what deleting frees.
Rust + GPUI. Inspired by the recent syntax.fm episode.
github.com/henrydennis/petal
1
36
A must-watch
Super excited to be interviewing @poteto, live, in ~48 hours on YouTube.
We'll be nerding out about skills, high-velocity software factories, and learning SOTA techniques for shipping with agents.
This Friday - 9AM PT. Don't miss it!
youtube.com/live/MN9dGgmLyso
1
20
Suat Özkaya | AI & Mobile retweeted
MCP 2026-07-28 says a server must protect requestState integrity when it influences authorization, resource access, or business logic. It doesn't say where the keys live, how twelve stateless replicas agree on them, or what enforces "use this confirmation at most once."
@mkoerbi follow-up walks through what the Python SDK covers by default and the three operational gaps it leaves open:
- Shared key distribution across replicas
- Key custody outside application memory
- Single-use enforcement, which is a state problem no amount of cryptography on a stateless token solves
The post pairs each gap with an OpenBao integration and a working example.
bit.ly/4Aw0Ee3















