Pranto C. | AI & SaaS Builder

You know what's the most annoying part of using both Codex and Claude Code? You end up becoming a human message bus between your own AI agents. I'll get a write up from GPT Sol, pass it over to Claude Opus 5.5 to work on the graphics, ask Codex to make changes, then bring the feedback back to the other agent. And God forbid one of them needs context from the previous conversation. Now I'm digging through chats, copying files, forwarding screenshots and explaining the same thing all over again. At some point you realize you're spending way too much time coordinating tools that were supposed to save you time in the first place. That's why Artil Connect caught my attention. In their demo, Codex directly messages Claude Code to review a PR. You can reach the same agent through email, Telegram or Slack, even when it's running on another machine. But there's an engineering problem worth thinking about here. (0__0) Agents can send messages back and forth all day. But if they're missing context, making assumptions, or can't verify each other's work, you've just automated the confusion. Say Codex has finished a feature and you want Claude to review it. Here's how I'd handle that handoff: - Tell Claude exactly what to review, like checking PR #42 for authentication and session handling issues. - Share the branch, changed files and any architectural decisions already made. - Include test results, known failures and anything Codex couldn't resolve. - Ask Claude to return specific findings with file references and suggested fixes. - Let Codex address the feedback, but keep the final merge approval with me. Now you've got a workflow where Codex implements, Claude reviews, Codex fixes, and you approve. You can start with just two agents and a well defined handoff. There's no point having five agents messaging each other if nobody can tell whether the work is actually finished. Check them out: artil.ai/connect

If you think your technical skills are becoming useless because AI can write code, I think you're seriously underestimating what's coming. I've been reading how engineers at top AI companies are working with coding agents. And honestly, it's making me rethink where I should invest my time as an engineer. Here are 5 things I'm personally focusing on: 1. Getting much better at codebase design. If you can read a 17,000-line file and reduce the input tokens needed for the same AI coding task by 83%. That's fascinating because good architecture now directly affects how efficiently your agents can work. 2. Developing deeper domain expertise, an agent can write a payment integration, but knowing how settlement, refunds, retries and reconciliation should behave is a different matter. You want to understand the business well enough to catch mistakes that perfectly valid code can still make. 3. LGetting better results from smaller models. I know Opus 5.5 makes everyone feel powerful, but here's a more interesting challenge. Can you get a cheaper model like Terra or Haiku to complete the same task reliably? Learning how to achieve that consistently is a skill I'd love to get better at. 4. Making my systems easier for agents to inspect. OpenAI gave its coding agents access to running applications, browser interactions, logs and traces. That way, agents could actually reproduce bugs and verify their fixes. I'm particularly interested in building more of this into my own workflows. 5. Understanding where the next engineering bottleneck appears. Anthropic went from roughly 2,500 to over 30,000 code changes merged daily. Suddenly, source control, builds and CI became major problems. I think there's a huge opportunity for engineers who can identify and solve these problems as AI increases development speed. The more I read about how these teams are working, the more I realize how much there is for experienced engineers to explore. Personally, I don't want to spend the next few years just getting faster at prompting AI to write code. We need to understand these systems well enough to make a smaller team capable of doing work that previously needed dozens of engineers. I'm considering writing a detailed article on these ideas, with real engineering examples and things you can try yourself. Would you be interested? Let me know in the comments.