Field notes on Cursor, Codex, Copilot, T3 Code, and the CI/review gap that still makes parallel agents feel like unpaid work.
Tooling changes every day. This is a July 2026 snapshot, not a ranking. I'm trying stuff in public because otherwise you fall behind.
For personal work I keep coming back to Cursor: Desktop, Agent tab, IDE, and Mobile. The Agent tab is good for juggling sessions. Some in local worktrees, some in the cloud, without losing track on one machine.
Mobile is very, very good. Not "fine for a phone." Actually good. I can spin an isolated agent while I'm jogging, open issues, nudge reviews, fix feedback, and I don't need my Mac half-open so some remote-control session stays alive. That Mac-must-stay-awake pattern (Claude / Codex remote control vibes) is a dealbreaker most days. Cursor's cloud + project + phone loop fits how I move.
Cloud agents are the other half. Each one gets its own computer, a real VM, so it can run the app, click around, and record verification videos / screenshots instead of handing you a blind diff. Less babysitting a local browser. Still not the CI/review sidebar I want (below), but better than "trust the model, merge later."
Then the glue I use every day: GitHub merge + preview deployments. I like it a lot. PR, merge readiness, and the Vercel (or whoever) preview in the same loop as the agent session. Boring product work, but it makes agents feel shippable instead of theatrical. Same with extensions. The wiring is better than the others right now. Less "install and hope," more "this actually connects."
At EnBW we don't use Cursor on the corporate stack, so work looks different. Warp as the hub, then OpenCode, Codex CLI, or Claude CLI depending on the job, worktrees in the terminal. Parallelism without the personal IDE.
I'm also spending more time in the GitHub Copilot app for that EnBW / GitHub-native world: issues and PRs in one place, local agents, worktrees, nested sessions, cloud runs. Model-agnostic and already where corporate code lives. That's a real advantage. T3 Code is the open-source itch: cool harness, and the browser / computer-use automation is neat. Claude Desktop? Barely touched it. Not enough reps for a strong take.
Every tool has a chat pane now. Fine. What's missing is lifecycle state.
I babysit PRs less because I love process and more because CI awareness is weak. Cursor already nails a lot of the shipping loop (merge readiness and preview deploys especially), but the status chrome still says "done" without the states that matter:
What I want is closer to a Conductor-shaped sidebar. I've only tried Conductor lightly, so treat it as an example of the shape, not a daily driver: hide sessions that need no human action (agent finished, CI still running), surface them again when action is required, and when approval lands, notify the agent and auto-merge. That loop is still too manual.
Run enough agents in parallel and you hit the other wall: you can't hold twenty threads in your head. Addy Osmani (and the Google I/O panel chat around it) called this the orchestration tax. Starting agents is cheap. Closing the loop still goes through one serial processor: you. Spawn too many and you're the bottleneck. Spawn one and you wait. Product UX has to own more of that middle state, or "agentic" just means fancier babysitting.
I'm reading Slow Productivity after Atomic Habits for the same reason I care about platform DX. The constraint isn't typing speed. It's attention, habits, and systems that don't fall apart under load. Agents don't remove that. They amplify whatever you already have, good or messy.
So yeah: trying everything, shipping with what fits the context, and waiting for tooling to grow up past sidebar chat.