Dmytro Durach

Board member at Onix, former CEO. Hands-on notes on AI agents and the tools I actually use

[Dmytro Durach @durach]

Don't make your Claude Code talk like a caveman. Make it report to you like a team lead: what's done and what you must decide.

[Dmytro Durach @durach]

How I use it, with the setup prompt: https://t.co/phBn2cmE85

[Dmytro Durach @durach]

A simple tool can work better than Jira for Claude Code. At least for me

[Dmytro Durach @durach]

@AgenticAIFdn From a security perspective, one agent = one virtual computer is probably the easier way to scale. I tend to build similar isolation in my agent infrastructure as well.

[Dmytro Durach @durach]

@simonw Well, crypto is still good for diversification. You should never put all your eggs in one basket.

[Dmytro Durach @durach]

@thorstenball Good advice. Sometimes I feel guilty when I ask an agent to redo the same work for the 10th time because I just changed my mind ;) Need to have a human mode and an agent mode

[Dmytro Durach @durach]

OpenAI created a skill for Codex to check PHP vulnerabilities. All of my potentially vulnerable sites were checked yesterday and fixed with Claude Code ;)

[OpenAI @OpenAI]

If you are wondering whether your website is affected, we have packaged a Codex skill to help check.

Point your Codex at this repo to guide you through it: https://t.co/XUEbOwXuQC

[Dmytro Durach @durach]

Correction to point 3: distros got the fix early, but not all have shipped it yet (remirepo: not today). Distro builds often keep the old version number, so check your distro's security notice for CVE-2026-109618 and CVE-2026-109619, not just php -v.

[Dmytro Durach @durach]

If any part of your infrastructure runs PHP, patch it today.

Two critical unauthenticated RCEs (CVSS 9.2), fixed in PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. mod_php is the highest risk.

https://t.co/66Wj6tgNnu

[Dmytro Durach @durach]

Can't upgrade tonight? Set session.upload_progress.enabled=0 in php.ini or the FPM pool (not ini_set) and restart every worker. Uploads keep working. It reduces exposure, it doesn't make you safe.

[Dmytro Durach @durach]

@emollick I'm surprised how much free time some people have

[Dmytro Durach @durach]

@GergelyOrosz Git log is still too useful to give up. It's how I look at the result

[Dmytro Durach @durach]

@SasuRobert Long answers frustrate me too, but what frustrates me more is not knowing what I should do next. That's why I changed my CLAUDE.md to keep it simple: https://t.co/zgNaLNsyPG

[Dmytro Durach @durach]

@aymannadeem What frustrates me: I can't thread a discussion with my agent in a terminal. One discussion splits into a graph, not even a tree. It's the same with people, only much faster. I believe we'll need a new kind of protocol for talking to agents

[Dmytro Durach @durach]

@donnfelker From what I see around me, developers work in terminals and chats, then hand tasks to agent factories (Hermes, OpenClaw, something else). Non-developers give tasks to the same factories through Kanban boards, Slack, Matrix, or other chats

[Dmytro Durach @durach]

@cbarmorecpa I think the key challenge will be security, as working with financial and PII data is not a trivial matter, especially when working with AI.

[Dmytro Durach @durach]

@weswinder I completely support you. This is why I've changed my main CLAUDE.md https://t.co/3kmlErt99s

[Dmytro Durach @durach]

@dexhorthy Still moving my pipeline to Hermes and haven't played with dot

[Dmytro Durach @durach]

@rauchg Why will? I think it does now.

[Dmytro Durach @durach]

The rule: https://t.co/zgNaLNsyPG

[Dmytro Durach @durach]

Don't make your Claude Code talk like a caveman. Make it report to you like a team lead: what's done and what you must decide.