Agents that run unattended, and leave something behind.

Scheduled agents are the easiest win in the whole category. A nightly dependency review, a weekly digest of what shipped, a run that watches an error feed and writes up anything unusual — none of that needs you awake, and all of it is work you'd otherwise never get to.

The awkward part is the output. It goes to a log file in a container that gets recycled, or an email that arrives at 3:04am and is read at 9:30 along with forty others, or standard output that scrolls past in a CI job nobody opens unless it's red. The agent did the work. The finding just never reached anyone.

And there's a second failure, quieter and more annoying: every run starts from nothing. Last night's agent found the same three deprecations tonight's agent will find, because tonight's agent has no idea last night happened. So the report is identical every day, which is the fastest way to teach a person to stop reading it.

Write to a folder, not a log

Point the run at a collection in your workspace — Reports/, say — with a _schema.md fixing what a report looks like: the date, what it covers, a severity, a status, a one-line summary. Each run appends a record; the body holds the detail.

That single change fixes the receiving problem. The output is somewhere permanent, at a stable path, in a format you and your agents can both read. Nothing is racing, either — each run writes its own record, so a weekly job and a nightly job that overlap simply don't interact.

The next run reads the last one

The better payoff is continuity. Because the reports are typed, tonight's run can start by asking what the last one said — the most recent record, or every open finding — and then report the difference: what's new, what's resolved, what's been sitting open for eleven days.

That's the line between a cron job and something that behaves like a colleague. One tells you the same thing every morning. The other tells you what changed, and stops mentioning the thing you already acknowledged.

Then you read it by asking

Once the reports are records rather than prose dumps, catching up stops being a scroll. Ask for everything above medium severity still open, newest first, and you get four records instead of a fortnight of logs. Ask your morning agent to summarize the week and it queries the same collection you would.

The reports also become context for regular work. An agent about to change the queue consumer can check whether the nightly run has been flagging something there for a week. That's a connection nobody makes when the reports live in an inbox.

Markbase isn't the scheduler

Worth being plain about: Markbase doesn't run your agents, trigger them, retry them, or notify you when one fails. There's no cron in it. The scheduling belongs to whatever you already use — a CI schedule, a timer, your harness's own runner.

What Markbase is, is the destination those runs can rely on: always there, reachable from wherever the job executes, versioned, and readable by every other agent you run. The unglamorous half, which happens to be the half that's genuinely hard to keep alive yourself.