Skip to content
vavio

Can one scheduled Hermes job use another job’s result?

Yes. In Hermes Agent, a scheduled job can list other jobs in context_from. Each time it runs, Hermes places the most recent saved output of each listed job above its prompt. It does not wait for an upstream job that is still running, and it does not check how old that output is. So schedule the collecting job well before the job that uses its result. A job can also read its own previous output with continuity.

This follows the official Hermes v0.23.1 documentation and source, commit 3b42157, checked on 15 November 2026. We have not run a job chain, timed one or tested its delivery on a Vavio computer. The examples explain upstream Hermes behaviour, not a ready-made workflow.

When is a chain worth setting up?

Split a job when collecting and judging are different kinds of work. Keep one job when the same run can do both.

Situation Documented approach
One job gathers data; a later job ranks, summarizes or formats it Give the later job context_from with the gathering job’s ID.
A daily digest combines several collectors List several job IDs in context_from.
A recurring scout keeps reporting the same items Turn on continuity so each run sees its own last report.
Every step must finish in order before the next starts Neither feature enforces that; see the timing section below.

A chain also lets each step have its own schedule, skills or model. If a collection step needs no judgment, it may not need the model at all: a script-only job or a pre-check can produce the data instead.

What does the second job actually receive?

The receiving job gets a block of text above its own prompt, headed with the upstream job’s ID. What the block contains depends on how the upstream run ended:

  • A successful agent run passes only its final reply, not its conversation, tool calls or files.
  • A script-only job passes its whole saved run file: a short header with the run time, then the script’s output.
  • A failed or blocked run passes its whole failure file: the header, the upstream job’s own prompt and the error.

The scheduler takes the most recently written run file in the upstream job’s output folder, under ~/.hermes/cron/output/<job_id>/, with these rules:

  • Quiet runs are skipped. [SILENT] replies, unchanged monitor ticks, script-only runs with empty output and pre-check runs that did not wake the agent are passed over, so an older real result is used instead.
  • An empty agent reply is not skipped. Hermes saves it as “(No response generated)”, and that text becomes the next job’s input.
  • Failed runs are not skipped. The guide keeps error documents eligible so the next step has recovery context. If the collector failed last, the second job receives that failure file as its input.
  • Long output is shortened. Text over 8,000 characters keeps its beginning and end, with the middle marked as omitted.
  • Missing output is silent. If there is no usable output, the block is left out and the job runs its prompt alone.

That last point matters. The receiving job is not told “the upstream result was missing.” If your summary depends on the data, write the prompt so it says what to do without it, for example: “If no collector output appears above, report that the collection step produced nothing.”

A results file the collector writes for itself, such as raw.md in the guide’s example, is separate. context_from passes the collector’s final reply. If the second job also needs the file, its prompt has to name the file and read it.

Does the second job wait for the first one?

No. The guide says chaining reads the most recent completed output and does not wait for upstream jobs running in the same tick.

If both jobs are due at 07:00, the second can start while the collector is still working, and it receives yesterday’s result. The guide’s example spaces the steps 30 minutes apart: collect at 07:00, rank at 07:30, publish at 08:00. Choose a gap longer than the slowest run you expect.

There is also no freshness check. If the collector is paused, broken or blocked on credentials for a week, the second job keeps receiving its last good or failed output. A successful agent reply arrives without the upstream run’s time, so ask the collector to begin its reply with the date and time it gathered the data. The receiving prompt can then say: “If the collected data is not from today, report it as stale instead of summarizing it.”

How do I reference the other job?

Use the upstream job’s ID, which cronjob action="list" or hermes cron list shows. One section of the pinned guide says names also work. In the v0.23.1 source, both the cron tool and the dashboard check each reference by exact job ID and refuse an unknown one as “not found”. The scheduler also skips any reference that is not a valid ID. An ID is the reliable choice.

The receiving job must be an ordinary agent job. A script-only job runs its script before any prompt is built, so context_from and continuity set on it are accepted but have no effect. A script-only job can still be the upstream step.

The simplest way to set it up is to ask Hermes in chat, which uses its cron tool. For example:

Create a job called “News collector” that runs at 07:00 and lists the five most relevant AI agent stories with links. Start the reply with the date and time collected. Then create “Morning brief” at 07:30 that uses the collector’s output through context_from and writes a three-point summary. If the collector output is missing or not from today, say so instead.

After creating them, ask Hermes to list both jobs. Its cron tool’s listing shows the second job’s context_from entry, which should be the first job’s ID.

Some changes break a chain without an error:

  • Removing the upstream job deletes its saved output. The receiving job still runs, now with no input.
  • Hermes keeps only the newest 50 run files per job by default (the cron.output_retention setting). Quiet runs count. A monitor that checks every few minutes and stays quiet for a long period can push its last real report out of the folder.
  • Recreating a job gives it a new ID. Update every job that pointed at the old one.

What does continuity add?

continuity makes a job read its own most recent substantive output before each run. Hermes frames it as the previous report, with an instruction to avoid repeating what was already reported and continue where the last run left off.

On its first run there is no earlier output, so the prompt runs as written. Quiet ticks are skipped as described above, so a quiet period keeps the last real report available. From the command line, hermes cron create "every 6h" "Scan for news" --continuity creates such a job, and hermes cron edit <job_id> --continuity or --no-continuity changes an existing one. In v0.23.1 we found no command-line flag for pointing at another job; ask Hermes to set context_from through its cron tool, or use the dashboard’s cron editor.

Continuity and an upstream source can be combined. A digest can read its collectors and its own last digest, so it reports only new items.

Continuity is not a database. It carries one earlier reply, shortened if long. For a few values that must persist, such as the last item reported, v0.23.1 also has a small per-job notepad: Hermes adds it to the job’s prompt on every run, and hermes cron notepad <job_id> set <key> <value> writes to it. A job that must remember every item it has ever seen should keep its own state file, and its prompt or script should read and update that file.

Is this the same as conversation memory?

No. Each scheduled run starts a fresh agent session. context_from and continuity inject saved run output into the prompt; they do not resume a chat. For what carries between interactive chats, see what Hermes remembers after a new chat.

Where should I start?

Start with two jobs whose output you can read. Run the collector once, check its saved result, then let the second job run at its scheduled time and read what it produced. Confirm it used today’s data and handled a missing input the way your prompt said. Only then add a third step or a delivery target.

The chain runs only while the Hermes scheduler is running. Does closing the terminal stop Hermes? explains what has to stay up, and why a scheduled job runs at the wrong time covers time zones, which matter if the steps use different ones.

Vavio provides a managed remote computer for Hermes at $14.99 per month, with model usage billed separately, so scheduled steps can run while your laptop is off. You still set up and check the jobs yourself. Compare the current hosting offer, or use the same upstream controls on your own server if you prefer to manage it.

Sources

Written by Vavio with AI assistance and checked against pinned official Hermes v0.23.1 documentation and source. This is a documentation and source review, not a pipeline, timing or delivery trial.

More from the Vavio blog: Hermes Agent guides