A quick snapshot of where coding agents are in 2026: you can hand one a real, multi-part task, walk away, and come back to merged PRs, a live deploy, and a few issues it worked around on its own.
Refactor to monorepo
Late at night, just before bed, I handed Claude Code a task in ultrawork mode:
Create a workflow:
- Refactor chmonitor.dev into a monorepo and add a landing page
from the design: https://api.anthropic.com/v1/design/h/xxxxxx?open_file=chmonitor+Landing.html
- Move the current main app from https://chmonitor.dev to https://dash.chmonitor.dev.
- Rename apps/web to apps/dashboard.
- Support cloud.chmonitor.dev but redirect to dash.chmonitor.dev.
- Deploy the landing page at https://chmonitor.dev (new: apps/landing).
- Rename apps/mcp-worker to apps/mcp (standardize).
- Docker: keep release code of the main app only + docs.
- Make as many PRs as possible with clear, scoped changes. Use /github:babysit-pr to auto-merge them one by one.
This was a big one. After a few rounds of exploration and an interview, it came back with a clear target:
I didn't have a strong read on the plan. I noticed I had only 30% left in my 5-hour usage window, and it would reset around 3 AM. So I told it to keep working and to find a way to wake itself up if it ran out of tokens mid-task. I went to bed.
It got through the refactoring workflows and built the new landing page, then moved and redeployed everything (mcp, docs, etc). At 00:12, since the local ScheduleWakeup caps at one hour, it pre-scheduled a Remote Cloud Agent to continue with the current context. It also flagged a concern: because this wasn't a local wake-up, some memory and context would be lost. I figured the local session could auto-recover once the usage limit reset.
It was a very long conversation by the next morning. The session limit was hit three times, which I think maps to three one-hour-capped local wake-ups. The remote trigger did fire, but the local session also recovered. It noted that the remote trigger had run, so it was checking status by pulling the PRs.
The one snag was the https://dash.chmonitor.dev deploy, but it found a way to recover. It left two hand-off items for me - which I'm going to pass to a fresh Claude session anyway 🗿.
Everything just worked, and as always I never looked at the code. It all deployed correctly:
That is roughly the state of coding agents in mid-2026: long-running and mostly autonomous. The interesting question is no longer "can it write the code" but how it manages itself over hours — planning, splitting work into PRs, merging them, and routing around its own token limits with scheduled wake-ups and cloud agents. I reviewed the whole run from my phone and never opened the editor.
What stands out to me is the harness, not just the model. I spent a good while in Codex, but Claude Code still does the most of any of them for this kind of long, self-managing run. The harness is good enough that it holds up even when I swap in a different model like GLM 5.1. That is the part that feels ahead right now.
