Today I spent the whole day on Codex automation.

My biggest impression:
The most valuable part of the agent is not
automatically doing everything for you.
But continuously pushing a bunch of stuck in-progress tasks
to the next clear checkpoint.
Today it did several types of tasks:
1. For those that can write code, directly fix in a clean worktree.
For example, a bug caused by misjudgment in a wiki compiler,
fix it in a clean worktree first,
after verification,
then cherry-pick back to the main repository with my approval.
2. For those that can be verified, directly refresh evidence.
For example, strategy preflight, watch artifacts, dashboard,
it will re-verify old blockers,
then write back to the current gate.
3. For decisions that require my approval, generate a decision card.
For example, whether to submit KYC,
whether to publicly follow up on affiliate issues,
whether to clear a dirty repo window.
This is very critical.
A good agent workflow,
the core is not “full automation.”
The core is:
it continuously advances low-risk tasks;
it stops for high-risk ones,
compressing decisions into a single sentence I can approve or reject.
Today I also hit a few pitfalls:
Dirty worktrees can block integration;
long logs can consume all context;
repeatedly refreshing old blockers wastes tokens;
stale locks can be left if the runner exits uncleanly.
So I finally changed it into two layers:
Resident layer:
lightweight scan every 15 minutes,
only look at new artifacts, locks, decision queues, task timestamps.
Deep push layer:
only when a real trigger occurs,
select a task to push to commit, report, parser, decision card, or waiting sample.
Running this setup,
feels more like a junior PM + engineer.
It doesn’t replace my product judgment,
but it continuously clears backlog,
turns fuzzy issues into clear evidence,
turns implicit bottlenecks into explicit decisions.
I’m increasingly convinced now:
the core ability of future personal workflows,
is not “whether you can code with AI.”
But whether you can design your task system,
in a way that an agent can continuously push forward.