AI Agent App — Reliability Engineer who works inside Claude Code + Replit

AI Agent App — Reliability Engineer who works *inside* Claude Code + Replit (diagnostic, not band-aid) Type: Ongoing, part-time to start (small scoped trial first) Rate: Tell us your rate in your proposal and why you're worth it Location: Open to remote / overseas Read this whole post before applying We get a lot of applicants who say "yes, send me the code, I can do it." That's not who we're looking for, and proposals like that will be ignored. We're looking for one person who thinks top-down, understands *why* a system behaves the way it does, and can tell us what's actually going on before touching anything. If that's you, the screening questions at the bottom are easy. If it's not, they'll be obvious. What we're building We have a functioning AI-agent application — think of it as an intelligent "chief of staff" that delegates work to sub-agents, uses tools and skills, and pulls in a user's files, calendar, and call/email context. It's live and adding real value for real users today. Stack is React + Node, running on Replit. We are not rebuilding from scratch and we are not looking to migrate off Replit right now. We're past the "this is broken, fix it" stage. The work now is reliability and architecture cleanup fine-tuning behavior, closing logic gaps, and untangling pathways that got built on top of each other over time. A concrete example of the kind of problem you'd own: an agent says "Great, I'll go do that for you," delegates the task to a sub-agent, and then… nothing surfaces. The user is left staring at a stalled response. Sometimes the work actually completed but never got surfaced back; sometimes the delegation logic itself loops or dies silently. Diagnosing *which* of those is happening — and why — is the job. The workflow (this is non-negotiable and the whole point) You will not be committing or merging code into our live app. Ever. Here's how we work: 1. We hand you a problem (often with screenshots / a description of the unexpected behavior). 2. You work the problem in Claude Code — start in plan mode to produce a clear task spec and your read on what's actually going on (this also keeps cost down). 3. You build the change in Replit on a separate branch / background build — never merged, never published. 4. You hand it back with: your diagnosis, what you changed, the expected result, and how to test it in the dev window. 5. We review, ask questions, and decide what (if anything) ships. If you don't already understand the difference between Claude Code plan mode, building on a background branch, and the Replit↔GitHub sync quirks that come from running an app out of Replit rather than out of GitHub — this role isn't a fit, and that's fine. You're the right person if you - Are a real engineer, not a "vibe coder." You can read a large, messy codebase and explain the logic flow top-down. - Are fluent in Claude Code (plan mode, branches, background builds) and comfortable in Replit and GitHub. - Know React and Node well enough to trace why a request goes through three places it shouldn't. - Default to "let me tell you what's actually going on" over "let me slap a fix on it." We do not want band-aids. - Are comfortable working on sub-agent / tool-using / delegation architectures and reasoning about where a hand-off can silently fail. - Move quickly, communicate clearly in writing, and don't need to be watched on every click. You're not the right fit if you - Want to migrate everything off Replit or rewrite it your way. - Plan to point an AI at the error and merge whatever comes back. - Need fully spec'd tickets and can't reason about intent on your own. Trial We'll start with a small, scoped, paid task so we can both see how you think. If it goes well, there's steady ongoing work. How to apply (this is the screen — follow it exactly) Applications that skip any of these go straight to the bin: 1. Start your proposal with the word `RELIABILITY` as the very first word, so we know you read this. 2. In 3–5 sentences, answer: *An agent tells the user "I'll go do that now," then the response stalls and nothing appears. Walk us through how you'd figure out whether the work completed-but-didn't-surface vs. the delegation itself failed — without changing any code yet. 3. Tell us, briefly, your hands-on experience with Claude Code specifically (not just "AI tools") and with Replit. One real example of each. 4. Confirm you're comfortable working on a branch you never merge, handing changes back for review. 5. Your rate and rough weekly availability. Short, sharp, and specific beats long. We're reading for how you think.