Skip to main content
Build and run

Building and Maintaining Automations from Your AI Assistant

There are three ways to build an automation in StackJack: the advanced builder in the portal, the chat assistant, and your own connected AI assistant through StackJack's MCP tools. They now build and…

Written By Christopher Scaminaci

Last updated 27 days ago

There are three ways to build an automation in StackJack: the advanced builder in the portal, the chat assistant, and your own connected AI assistant through StackJack's MCP tools. They now build and maintain the same things. Anything the portal builder can set, the other two can set too.

From your connected AI assistant

For the recommended way to build, including a downloadable builder skill that interviews you, builds and tests the automation, and example requests, see Build Automations with Your AI Assistant. This page is the tool-by-tool reference.

Editing a live automation, not only a draft

stackjack_update_agent_draft keeps its name, and it now edits a live automation as well as an unfinished draft. You no longer have to stage a copy, test it, and deploy it over the original just to change a schedule.

Four transitions are still handled by one tool each, so two tools can never disagree about the same switch:

What you want to changeThe tool that does it
Activate, deactivate, promote out of dry run, demote, resync, re-sync every automation (resync_all)stackjack_set_agent_lifecycle
Accept the destructive-action or production acknowledgmentstackjack_accept_agent_consent
Everything else about the configurationstackjack_update_agent_draft

If you send an activation or dry-run change to the update tool, it refuses and names the tool that owns it.

triggerConfig and the tool set are replaced wholesale, not merged. When you send either one, send it complete. Changing a cron and leaving out the smart filter you had removes the filter.

The automation type, the definition of done and the spending limit

stackjack_draft_agent and stackjack_update_agent_draft take three more settings. All are optional.

SettingWhat it sets
kindstandard or outcome. Omit it for a standard automation; on an update, omit it to leave the type unchanged.
outcomeThe definition of done of an outcome-based automation: a checklist, a result file name and the number of attempts. It is required when kind is outcome. On an update, send an empty string to clear it and omit it to leave it unchanged.
maxUsdPerRunThe spending limit per run in US dollars, for runs on your own Anthropic key or your Azure Foundry. Omit it for no limit. On an update, omit it to leave the limit unchanged; 0 or less removes it. With StackJack credits the limit is maxCreditsPerRun.

An outcome-based automation is graded by a checker against a definition of done, and the checker can send it back to redo its work. When a save changes the type to outcome-based, the answer carries this warning, and your assistant shows it to you:

The checker can send the automation back to redo its work. A redo can repeat changes it already made in your systems.

Read the definition of done with stackjack_get_agent_definition, section outcome; the answer also says the automation's type, its dollar limit and whether the limit counts US dollars or StackJack credits. Write the whole definition of done with stackjack_set_agent_definition, section outcome. stackjack_describe_agent_schema carries the author guidance: the checker reads only the result file, so each item should be something that file must show.

A save that breaks a rule is refused with a sentence your assistant shows you: an outcome-based automation cannot be combined with Split big jobs across parallel agents, it does not run on Azure Foundry, and outcome-based automations can be switched off for an organization.

All three ways to build can set a tool chain to the ordered level: the advanced builder, the guided wizard and your connected AI assistant. An out-of-order step in an ordered chain stops the run; see Guardrails and Safety.

Eight new tools for the parts that were portal-only

ToolWhat it is for
stackjack_get_agent_definitionRead one document at a time: the fixed start and finish steps, the tool chains, the trigger configuration, the tool policy, the version list, or the difference between two versions.
stackjack_set_agent_definitionWrite the fixed-steps document or the tool-chains document. It answers with what it accepted — size, checksum, whether an acknowledgment was cleared — and never echoes the document back.
stackjack_describe_agent_schemaThe authoring contract, with worked examples: step kinds including decision steps, the {{…}} value grammar, tool chains, tool policy modes, trigger configuration including a smart filter, and the decision question model. Read this before writing a document by hand.
stackjack_inspect_agentRecent activity: webhook deliveries, smart-filter judgements, a run's decision steps, lifecycle history, the schedule, a tool-trim preview, proposed tool grants, and runs.
stackjack_apply_agent_maintenanceRestore a version, apply a tool trim, re-register a schedule, rotate the webhook secret, archive, or accept a proposed grant. Every action echoes a phrase StackJack computes, so you confirm the specific action rather than a generic yes.
stackjack_control_agent_runCancel a run, or approve or deny a call it paused on.
stackjack_manage_agent_templatesList, read and install automation templates.
stackjack_manage_agent_memoryList, read, redact or delete an automation's memory.

stackjack_manage_agent_knowledge also gained three verbs: re-fetch a knowledge source, attach a file directly, and edit a source in place (a text source's name and text, a URL source's name and address, a file source's name only).

You will not see the new tools until your connection is re-granted. A connection whose grant was a hand-picked subset of tools keeps exactly the tools it was granted. Open the endpoint's grant page in the portal and grant the new ones, then refresh your assistant's tool list.

A person has to be behind the change

Every mutating call from an assistant is refused unless StackJack can resolve the signed-in person making it. A machine-only client — one created by an administrator with nobody bound to it — can read, and cannot change an automation at all. The person it resolves then passes exactly the same permission checks a portal save would apply, and the audit record shows both the person and the fact that a machine client carried the change.

What is deliberately held back

  • A rotate tells you it rotated and nothing else. The new webhook secrets are never returned to an assistant; read them in the portal.
  • A restore answers with the same curated summary any other write does, never the automation's raw stored record.
  • Approving or denying a paused call has no confirmation phrase. It is contained differently: the claim on that paused call can only be taken once, the approval is single-use, and the decision log records that a machine client answered it and which one. Approve from an assistant only if you would have approved it in the portal.

The acknowledgment is re-asked when the write steps change

If an automation has fixed steps that write to your systems, someone should acknowledge them. Adding, removing or moving a write step — or any step above one — or changing its tool withdraws that acknowledgment, on every lane, whether the change came from the portal, the chat assistant or your own assistant. Pointing one of its values at an option a decision step picked, or changing that decision's options, withdraws it too. Changing a write step's fixed value, or which value from the trigger it uses, does not. The response says so in plain words.

No lane refuses to deploy an automation without the acknowledgment: it is a warning. But a step whose tool StackJack marks as destructive does not run until someone acknowledges the new list, and on some connectors that mark covers creating or updating a record, such as adding a ticket note, not only deleting something.

A run that was queued before the write steps changed, and starts after the new list was acknowledged, does not carry out any of its steps that change something, destructive or not, because the approval covers the new list and the run follows the older one. The next run uses the new list.

Two things follow that are worth knowing:

  • Acknowledging is not a one-time formality. The phrase you confirm names how many destructive tools and how many write steps the acknowledgment covers, so it is always about the list in front of you.
  • An acknowledgment for an automation that has no write steps is not withdrawn by an edit that adds none.

From the chat assistant in the portal

Edit with the assistant on an automation's detail page opens the chat builder against that automation instead of a blank one. It loads the automation as it stands, you describe the change in conversation, and you finish in the builder's own Save — the same save the advanced builder uses, against the same automation.

  • Your knowledge survives the edit. Core knowledge and knowledge sources are loaded with the rest and saved back untouched unless you change them. If StackJack cannot read an automation's knowledge, the edit lane refuses to open rather than opening with it blank — the advanced builder refuses the same automation for the same reason.
  • Knowledge sources are added, changed or removed only when you say so. Asking for a new runbook or link adds it and leaves the others alone. Asking to change one changes that source in place and keeps it the same source: a text source can get a new name or new text, a link a new name or a new address (read again at once, on the same refresh schedule), and an uploaded file a new name only — to replace a file's contents, remove it and attach the new file in the advanced builder. A source is removed only when you ask for that source to be removed. If the assistant names a source StackJack cannot match to exactly one attached source, nothing is changed for it and the page tells you so.
  • The assistant can author fixed steps, including decision steps, through the same editor the advanced builder uses.
    • A decision step judges only what you hand it. Under What to look at it takes the values its questions are judged on, and it sees nothing else: the answers of the steps above it are not passed on their own. A decision step with nothing there would be asked about nothing, so StackJack refuses to save it and says what to add. Good things to hand it are one field of one item, such as a ticket's summary, or a short list from the trigger or an earlier step. Where a tool's answer carries a count, that works too: HaloPSA list tools return one beside their list, and other connectors do not, so for those hand the decision the list itself. Everything one decision step reads together may not be larger than 64 KiB. A bigger one is refused with a message naming the largest value, and it is never shortened, so ask the assistant to keep the list short.
    • A step that reads can stop the run when it finds nothing. Ask for "stop the run if there are no tickets" and the assistant adds a condition to the step, which StackJack checks with no AI call. A run stopped that way ends as Skipped, not Failed, and sends no failure email. See Stop when a step finds nothing.
  • Removing something has to be said. The assistant treats silence as "leave it alone", so say "remove the schedule" rather than describing the automation without one.
  • The destructive acknowledgment is yours, not the assistant's. The tick sits beside the conversation and the model can never set it for you.
  • Four more settings can be set in the conversation. Test mode, recording the automation's runs, requiring a passing evaluation before a staging copy is deployed over it, and skipping a repeated webhook delivery. The last three can be set while you create an automation and while you edit one. Test mode can be changed only while you edit, because a new automation always starts in test mode. When you edit, only the settings you ask to change are saved, so a change someone made on the automation page in the meantime is kept.
    • Test mode changes right after your changes are saved. Before you save, the panel says so in one line — that saving will also switch the automation live, or pause it to dry run. If StackJack cannot take the automation out of test mode, your other changes are still saved and the confirmation says why. If a knowledge source could not be attached, the automation is not taken live: fix the source and save again.
    • Recording keeps real tool calls and your customers' data, so the assistant asks with the same consent sentence the automation page shows, and recording is turned on only when you pick Yes on that question yourself. Typing "yes" is not enough. Until you pick Yes, the panel says that recording will stay off, before you save and again after. Turning recording off needs no question.

What no lane can do

One thing is missing from every lane, the portal included, and this is not an assistant limitation:

  • Evaluations. No lane browses fixtures, creates scenarios or launches evaluation runs.