Skip to content

EP Agent Jobs

EP Agent Jobs runs AI agents on a schedule, from your site, whether or not anyone is at a computer. You define a job — when it runs, what to ask, which model, what it is allowed to touch, where the answer goes, and what it may cost — and the site takes it from there.

It rides the EP Cron heartbeat, so it needs no visitor traffic to fire and nothing stops when you close your laptop.

Published by ElmsPark Studio.

Use cases:

  • A Monday morning digest. Summarise last week’s bookings or orders and email it before you sit down.
  • A watchdog. Check something daily and only speak up when it looks wrong.
  • A recurring content chore. Draft the thing you rewrite every month, ready for editing rather than starting blank.
  1. You create a job: schedule, timezone, prompt, model, allowed site actions, delivery channel and monthly budget.
  2. The EP Cron heartbeat wakes the scheduler and asks which jobs are due.
  3. Due jobs run, up to the per-heartbeat limit and the wall-clock limit per run.
  4. The result is delivered by email, Telegram or Buzz.
  5. The run is recorded, with history kept for as long as you choose.
  • PageMotor 0.11.5 or later
  • EP Cron, providing the heartbeat that drives the scheduler
  • EP Suite base class (bundled with the plugin)
  • An Anthropic API key, and credentials for whichever delivery channel you use
  1. Install and configure EP Cron first, and confirm its heartbeat is actually arriving. A scheduler with no heartbeat is silent rather than noisy — nothing runs and nothing complains.
  2. ep-agent-jobs.zip comes with an EP Suite licence — ElmsPark supplies it directly (see EP Suite plugins); after install it updates through your site’s Updates screen.
  3. Upload via Plugins → Manage Plugins. Activate.
  4. Open Plugin Settings → EP Agent Jobs.
SettingWhat it does
What site actions may an agent run?The ceiling on what any job can do to your site. Set this deliberately; it is the main safety control.
Default email recipientWhere results go when a job does not name someone else.
Telegram chat ID / Buzz channel, relay URL, roomCredentials for the non-email delivery channels.
Default timezone for new jobsSo “9am” means what you expect. Each job can override it.
Jobs per heartbeatHow many due jobs may start per beat. Keeps a backlog from stampeding.
Wall-clock limit per runCaps a single run that would otherwise hang.
Keep run history forHow long finished runs are retained.
Pause all agent jobsA single stop switch for every job at once.

The action ceiling is the setting that matters

Section titled “The action ceiling is the setting that matters”

“What site actions may an agent run?” decides what an agent is permitted to do, not merely what you meant it to do. A prompt is not a permission boundary: an agent that misreads its instructions is still limited by this setting and by nothing else. Start restrictive and widen only when a job genuinely needs it.

The job form offers ready-made jobs under Start from a template: a daily lead scout, an inbound lead digest, a weekly pipeline report, a competitor and market watch, a content radar and, from 0.1.13, a weekly landing page report. Picking one fills the form in; change anything before saving.

A template never ticks which site actions or web domains a job may reach. You choose those yourself.

Every Monday at 08:30 it emails a short report on each advert’s landing page over the last 7 days: views, leads, leads per 100 views, and up to three things worth looking at. It reads EP Lead Source (leads per landing page and per advert tag) and EP Analytics (page views per advert).

To use it:

  1. Turn on Site actions for the job and allow EP_Lead_Source:leads-by-source, EP_Lead_Source:list-leads and EP_Analytics:campaign-pages.
  2. Set What site actions may an agent run? to Admin. Those reports are admin-only, so a job cannot read them at a lower setting. Saving the job warns you if this is not done yet.
  3. In EP Analytics, switch on Adverts and campaigns so it counts views per advert.

If a number cannot be read or matched, the report says so and writes “not available” rather than guessing. It never includes anyone’s name or email address.

An assistant connected to your site (Claude through the site’s MCP connection, for example) can create a job for you with the create-job action, from a template or from scratch. Ask it in plain words, such as “set up the weekly landing page report”.

It is checked exactly as the job form checks it, so anything the form would refuse, the assistant is refused too, with the same message. On top of that:

  • A job an assistant creates starts paused, unless you told it to start the job. Nothing runs or is spent until you switch it on in the jobs list.
  • At most 20 jobs on the site, no schedule more often than hourly, and a monthly budget above $0 and no more than $50. Anything beyond that can still be set in the job form.
  • Only an admin connection can do it, and each job created this way is recorded in your site’s audit log.

Editing and deleting jobs stay in the job form.

Each job carries a monthly budget, and the settings page shows the month’s spend so far. A job that has spent its budget stops rather than continuing quietly, which is the behaviour you want from something that runs unattended.

Nothing ever runs. Almost always the heartbeat. Check EP Cron is active and its last heartbeat is recent; the plugin’s dashboard card reports a missing or stale pulse rather than assuming one.

Jobs run at the wrong time. Check the job’s timezone rather than the default, and remember that the schedule is evaluated when a heartbeat arrives, so a sparse heartbeat makes every job late.

A job stopped mid-month. Check its budget and the wall-clock limit before assuming a failure.

  • Changing this plugin’s settings over the API now needs an admin token. Producer tokens and AI apps connected through a producer account can still read the settings. Saving from the Admin screen works exactly as before.
  • A new starter template: Weekly landing page report. Every Monday morning it emails a short, plain-English report on each advert’s landing page: views, enquiries and sign-ups, leads per 100 views, and up to three things worth looking at. It reads EP Lead Source and EP Analytics on your own site.
  • It says when a number is missing. If a report cannot be read, or two figures cannot be matched, the email says so and writes “not available” instead of guessing. It never includes anyone’s name or email address.
  • To use it, allow the three site actions named in the template’s description, and set “What site actions may an agent run?” to Admin, since those reports are admin-only.
  • Your AI assistant can now create a job for you. Ask Claude (or any assistant connected to your site) to set one up, from a template or from scratch. It is checked exactly as the job form checks it, and refused for the same reasons.
  • Jobs an assistant creates start paused, so nothing runs or is spent until you switch them on, unless you told it to start the job. An assistant cannot create more than 20 jobs on a site, schedule one more often than hourly, or give one a monthly budget of more than $50 or no cap at all. Every job it creates is recorded in your site’s audit log.
  • Saving a job now warns you about site actions it could never call: one whose plugin is not installed, or one that needs a higher setting under “What site actions may an agent run?”. The job still saves.
  • Needs PageMotor 0.11.5 or later. Nothing to do after updating.
  • Jobs run on the Claude 5.5 models. The model menu offers Claude Sonnet 5.5 (the default), Opus 5.5 and Haiku 5.5. A job saved on Sonnet 5, Opus 5 or Haiku 4.5 runs on its 5.5 successor with nothing to re-save.
  • Opus 5.5 costs a fifth less than Opus 5, and Haiku 5.5 about a tenth of Haiku 4.5. The spend figures use the new prices.
  • Also carries 0.1.11: the jobs list shows a job’s failure streak and which tool calls failed.
  • Needs PageMotor 0.11.5 or later.
  • The jobs list now shows when a job keeps failing. A job whose scheduled runs have failed or been skipped since its last success carries a small red note beside its status, such as “failing 3 in a row”. It goes away as soon as a run succeeds.
  • Run history shows failed tool calls. When some of a run’s tool calls were refused, blocked or could not reach their site, the Tools column now says how many beside the total, for example “1 failed”.
  • The Command Centre tile counts these jobs too. “Needs a look” now includes a job that is failing or being skipped in a row, not only one whose last run errored or hit its budget.
  • Nothing to do after updating.
  • You are now told when a scheduled job keeps failing. If a job’s scheduled runs fail or are skipped twice in a row, you get one message on that job’s own delivery channel (email or Telegram) saying what went wrong and when it will try again. Until now a failing job stayed silent until someone opened the admin screen.
  • One message per problem, not one per run. While a job keeps failing the same way you hear nothing more. If it starts failing differently, you get one new message. The count starts again as soon as a run succeeds.
  • A failed Run now is not counted, since you are already looking at the result.
  • Digests now say when the agent could not look. If any of a run’s tool calls were refused, blocked or could not reach their site, the delivered result ends with one line saying how many, so “nothing found” and “could not check” no longer read the same.
  • Nothing to do after updating. The database update runs by itself.
  • A busy or rate-limited Anthropic no longer fails a run so easily. When Anthropic says it is overloaded, too busy or limiting requests, a run now tries up to three times instead of two, and waits as long as Anthropic asks before trying again instead of a fixed three seconds.
  • A run never waits past its own time limit. If Anthropic asks for a longer wait than the run has left, the run stops at once and its log says how long Anthropic asked to wait.
  • Every other error is worded as before.
  • Safer web fetching. When a job reads a web page, it now connects only to the address that passed the safety check, so a domain cannot switch to an address inside your own network halfway through.
  • A clear warning when agents may change your site. While “What site actions may an agent run?” is set to Producer or Admin, the settings screen explains the extra care needed for jobs that also read web pages.
  • Nothing to do after updating.
  • Security fix: your keys and tokens are now stored encrypted. The Anthropic key, Telegram bot token, Buzz key and crawl token are encrypted on your site. Ones you saved earlier are encrypted the next time you open the admin. Nothing else changes.
  • Settings language menu. The language menu in this plugin’s settings now lists only the languages it is actually translated into, plus English, so you can no longer pick a language that changes nothing.
  • Section status. Each settings section shows whether that feature is switched on. On a site that also runs an older EP plugin, the section shows its plain title instead.
  • Removes a function call that PHP 8.5 reports as deprecated, so it no longer fills your error log.
  • Corrects the PageMotor 0.11.3 preparation shipped in 0.1.4. That release grouped this plugin’s API and MCP actions into families, but in a shape PageMotor 0.11.3 does not accept. On 0.11.3 the plugin would have registered none of its actions, and nothing on screen would have said so.
  • This version uses the shape 0.11.3 expects, and keeps the earlier shape for sites still on an older PageMotor. One build serves both, so there is no order you have to update things in.
  • Safe to install now. Nothing you can see changes.
  • Prepares the plugin for PageMotor 0.11.3, which requires every plugin to group its API and MCP actions into named families. On that release an action declaring no family is refused silently, so without this update the plugin’s actions would stop being offered with nothing on screen to say why.
  • Families exist so that an AI working on your site sees the related actions together instead of stopping at the first one it tries.
  • Safe to install on your current PageMotor. No behaviour changed.

Fixes “Your session has expired. Please reload to ensure your security.” on PageMotor 0.11, which affected this plugin’s admin screens through the shared suite header’s language selector and brand-colour picker. Anonymous visitors were never affected.

PageMotor 0.11 hardening for the job action ceiling. Scheduled runs inject a token whose access tier comes from the action-ceiling setting, stored using the old read-only name. PageMotor 0.11 renamed that floor tier to open with no alias, and although 0.11 fails closed by coercing an unrecognised tier to open, jobs were working by coercion rather than by contract.

The dispatcher now reads the running core’s actual tier ladder and translates the stored name when the core uses the new vocabulary, rather than comparing version strings. Behaviour is identical on both cores and your stored settings are untouched.

Dashboard card readiness. The status pill now reads the real heartbeat timestamp rather than merely confirming the cron class loaded, so it shows red when no heartbeat has arrived and amber when one is stale.