Coding in 2025 (part ii)
Conductor + GPT-5-Codex: the 2025 coding loop goes real
Sep 24, 2025
Conductor.build + GPT-5-Codex finally feels like the workflow I described in Coding in 2025: parallel tries, fast feedback, diff-first review. Running the OpenAI Codex CLI inside Conductor workspaces is real, not a toy.
This is the “Coding in 2025” loop
Conductor is the first non-skeuomorphic IDE for the AI era. It treats work as threads, runs, and diffs, not files and tabs.
- Start multiple agent threads against one goal.
- Read traces like you read test output.
- Review diffs, not entire file trees.
- Merge the strongest path, discard the rest.
- Do it anywhere. The loop works from the back of an Uber.
GPT-5-Codex
🤯 GPT-5-Codex just changed my view on the future. I needed a beautifully animated video with rich content and visuals. This needed to be a react app for... reasons. No model could do it. Few could even get a working example. GPT-5-Codex crushed!! 🔥 >1hr of uninterrupted
GPT-5-Codex behaves like a dedicated coding agent. Paired with the Codex CLI, it keeps context tight around the repo and ships cleaner changes.
- Short prompts, high signal.
- Multi-file refactors land as readable PRs.
- Fast on simple tasks, digs in when problems get hard.
- Feels purpose-built for agentic coding, not chat with code on the side.
Parallel agents in a single codebase
This is where Conductor clicks. You get parallelism without chaos.
- Parallel agents, one codebase, all local, with live reload.
- Manage it like an engineering team: set priorities, acceptance tests, and guardrails.
- Code is authored for you: features implemented, tests added, docs updated.
- PRs are drafted for you: branches, commits, descriptions, checklists.
- Reviews are written for you: inline comments, summaries, suggested fixes.
- Review feedback is applied for you: agents pull comments and push follow-up commits until checks pass.
- Your job shifts to orchestration: decide what gets done and how it ships.
Parting thoughts
Parallel is the new default. Conductor + GPT-5-Codex makes the loop trivial: start multiple threads, read the traces, merge the best—your bottleneck is now habits, not tooling.
You have firepower; aim well.