SquidSupport
Product reference / AI

Squid Agent

In the app: AI → Squid Agent

Needs the 'Use Squid Agent' permission.

Squid Agent is an assistant for the people who run support. You ask it about your tickets, reports, setup and settings, and it answers with numbers that link to their source. It can also prepare changes to your setup, but it never makes one itself: every such change waits as a proposal until a person approves it. What it remembers for you and the skills you save yourself can be saved at once; see "Memory" and "Skills".

What Squid Agent is and who can use it

Three things are needed to use Squid Agent. They are checked in this order, and the first one that fails is the one the page names:

  • It is switched on for your organization. There is no switch for it in the app. Squid controls whether new organizations start with it on, and can switch it on for existing organizations. When it is off, Squid Agent is hidden from the menu and its page shows no content. If it is switched off while the page is open, a later request can show "Squid Agent isn't enabled for your organization".
  • Your role has the 'Use Squid Agent' permission. Owners and Admins have it by default. For other roles an admin ticks it under Settings → Roles. Without it, Squid Agent is hidden from the menu and its page shows no content. If your permission is removed while the page is open, a later request can show "You don't have access to Squid Agent".
  • Your organization has an active plan or trial. Without one, a banner says "Choose a plan to use Squid Agent" and new questions are off.

When the first two hold, the "AI" item in the main menu opens Squid Agent, and "Squid Agent" is the first row of the AI menu. The plan is not checked for the menu: without a plan the page opens and shows the banner.

Squid can pause Squid Agent, for everyone or for one organization. A banner then says "Squid Agent is paused by Squid". While it is paused you can still read threads, findings, proposals and activity, give feedback, and reject or cancel a proposal. You cannot ask a question, approve, retry or roll back a proposal, use "Approve selected" on memory suggestions, or start a simulation. An answer that is running when the pause starts stops at its next step and shows "Stopped by Squid". When Squid cannot read the pause switch, Squid Agent behaves as paused and says "Squid Agent is temporarily unavailable".

The page has ten tabs: "Mine", "All org", "Proposals", "Findings", "Incidents", "Goals", "Memory", "Skills", "Schedules" and "Activity". "All org" is only shown with the 'View Everyone's Squid Agent Threads' permission.

Squid Agent works with your own access. It only reads the stores and the reports you may see yourself, and a change it prepares can only be approved by someone who may make that change by hand.

Answers and the daily limit

The meter at the top right shows how many answers were used today against the limit, and when it resets. Every question that is accepted uses one answer, also when it is asked from the Squid Agent panel on another page or by a schedule. The limit is for the whole organization, not per person, and resets at 00:00 UTC.

The limit comes from your plan:

  • trial: 10 answers a day. An organization on a trial counts as a trial whatever plan it picked;
  • Starter: 25;
  • Basic: 50;
  • Pro: 100;
  • Advanced and Enterprise: 200;
  • any other plan: 25.

Squid can set a different limit for one organization; that number then replaces the plan's.

When the limit is reached, a banner says "Today's Squid Agent limit is reached" and a new question is refused until the reset. A refused question is not saved and not queued. A scheduled run that falls after the limit is skipped and marked "Skipped: daily answer limit reached". Reading threads, approving proposals and everything else on the page keeps working.

An answer is given back in two cases: the question ended without the AI being called at all, for example because you stopped it while it was still waiting; or it failed on an error on Squid's side before the AI had written or looked up anything. A question that was answered keeps its answer used, and so does one that was stopped or failed after the AI had started on it.

An organization that has not finished setting up also gets 100 answers that do not count toward the daily limit. They are used first for questions you send yourself, including from the panel, and end when they are used up or when the first store's setup is complete, whichever comes first. From then on every question counts toward the daily limit. Scheduled runs always use the daily limit, including during setup.

Besides the daily limit there are two allowances per person. By default, they hold up to 6 and 60 questions and refill continuously, at 6 a minute and 60 an hour. A question needs room in both. Squid can change these limits; during a service outage, a reduced allowance may apply. A question past it is refused with "Too many questions at once" and the number of seconds to wait; it is not sent or queued. One thread answers one question at a time: a second question in the same thread is refused with "Another answer in this thread is still running".

Threads

A thread is one conversation with Squid Agent: your questions and its answers, kept with the sources behind each number.

"Mine" lists your own threads, most recently used first, 50 at a time with "Load more". Each row shows the thread's name, when it was last used and the stores its answers read. "New thread" at the top right starts an empty one. With no thread selected, the right side is an empty new thread with the question box.

On a thread of your own you can "Rename" it (up to 200 characters), "Archive" it and "Delete" it. An archived thread moves to the "Archived" group at the bottom of the list, which opens with "Show"; "Unarchive" brings it back. Deleting asks "Delete this thread? This can't be undone." and cannot be undone.

"All org" lists everyone's threads. It needs the 'View Everyone's Squid Agent Threads' permission. Until you select a thread, the right side says "Pick a thread". Other people's threads are read-only and marked "Read only": you cannot ask in them, rename, archive or delete them, or rate their answers. A question and its answer in someone else's thread are hidden from you when the answer used a store you may not see or a report your role may not open; the row then says "Answer hidden" and shows only who asked and when. The Owner sees every finished answer. An answer that is still being written shows "In progress" to everyone except the person who asked, the Owner included.

Threads and their answers are kept. The step-by-step record of how an answer was made is removed 180 days after the question; the answer itself, its numbers and its sources stay readable. A thread in which nothing was ever asked is removed after a day.

Asking a question

Type in the box at the bottom and press Enter or "Send". Shift+Enter starts a new line. A question can be up to 4,000 characters. Squid Agent answers in the language you write in.

While it works you see the steps it takes. "Stop" cancels the answer. An answer that takes longer than five minutes is stopped by Squid and says "Stopped because the answer took too long"; a question you ask yourself that prepares a change to your AI configuration gets fifteen minutes.

A question that could not be sent stays in the thread marked "Not sent", with "Send again". A failed answer says "This answer couldn't be finished". It has no retry button: ask the question again, which counts as a new question.

What an answer can contain

  • Numbers. Every number can be opened as "Where this number comes from": the period, the stores, the filters, the source, the definition, how fresh the data is and whether it was read from your reports, read from your settings, calculated by Squid or estimated. Where a number counts tickets, "View tickets" lists the tickets behind it and opens them in the Inbox; it needs the 'View Reports' permission and access to every store the number counts.
  • Tables, charts and ticket lists. A chart's bars can be clicked for their source, and "Show the values" lists them as text.
  • Setup cards. A checklist of what is set up, a next step with a button that opens the page for it, the DNS records to add for a domain with "Copy" buttons, and a connection check. When the check found a problem, "Check the connection again" sends that as a new question in the thread: Squid Agent reads the connection once more and changes nothing, and it uses one answer.
  • "Squid support request". When Squid Agent cannot solve something, it writes a summary for Squid's support team. "Copy summary" and "Copy address" let you mail it yourself. Nothing is sent automatically.
  • "Coming later". What Squid Agent would propose once it can make that kind of change. It is for information only.
  • A proposal card, with "Approve" and "Reject". See "Proposals".
  • A memory. "Squid remembered: …" with "Undo" appears when something was saved at once. "Remember this?" is a suggestion that needs a click. A grey "From memory" chip in the text shows which memory an answer used; open it for "Why does Squid know this?". See "Memory".
  • "Save as finding?" appears when Squid Agent reached a conclusion in a conversation that read customer-written text. It does not save the finding itself then; your click saves it, with your access. It needs 'View Reports', and saving an impact estimate on an existing finding also needs the permission that changes that finding.
  • "Decision needed" is a QA question: failed QA checks point at a rule your policy does not cover. Pick one of the options and click "Save decision". Seeing the card needs 'View QA Reports'; saving needs 'Edit QA Settings'. Saving creates the follow-up proposals, each of which is approved separately.
  • "Save this as a skill?" appears under an answer you just got when the same kind of look-up succeeded twice within 30 days. See "Skills".

Under an answer

  • "Helpful?" lets you rate the answer "Helpful" or "Not helpful". "Not helpful" opens a comment box of up to 2,000 characters. Only the person who asked can rate. The rating shows in the "Activity" tab.
  • "Save as skill" is shown under an answer of yours that looked something up. It keeps look-ups this answer made, so you can run them again by name. See "Skills".

"Notes" is a button at the top of the thread, not under an answer. It opens what Squid Agent wrote down for this conversation only. "Clear notes" removes them and is for the person whose thread it is. Notes are not memory and stay in the thread. After a conversation reads customer-written text, Squid Agent writes no new notes in it. Earlier notes remain until cleared.

Proposals

A proposal is a change Squid Agent prepared, for example to your AI configuration, a knowledge item, an automation rule, widget FAQ entries, product knowledge, the QA rubric, an A/B test, a goal, a schedule, a shared skill or an organization memory. Nothing changes until a person clicks "Approve".

The tab lists proposals with their number, kind, stores, who asked, who approved, status and date. The filter shows "Pending", "Approved", "Failed", "Rejected", "Expired & stale", "Cancelled" or "All". With the 'View Everyone's Squid Agent Threads' permission you can switch between "Asked by me" and "Everyone"; without it you see the proposals you asked for. A proposal about a store or area you may not see is listed as "A change outside your access".

Click a proposal to open its thread at the card. A proposal without a thread opens on its own page, and someone else's pending suggestion to remember something opens Memory → "Pending". Things to know:

  • Who may approve. If you asked for the change and you could make that change yourself when you asked, in every store it touches, only you approve it. Otherwise anyone with 'Use Squid Agent' who has the needed permissions and stores can. A schedule and a personal memory are approved by their owner only. The card names who can approve and why you cannot, when you cannot.
  • It runs as the approver. The change is made with the approver's access and is recorded with their name. Their access is checked again before every step; if they lost it meanwhile, the change stops.
  • One by one. There is no select-all and no bulk approve for changes.
  • "Reject" needs a reason of up to 500 characters and is for people who could approve, except the person who asked. The person who asked withdraws their own proposal with "Cancel request".
  • Expiry. A pending proposal expires after 7 days; a suggestion to remember something after 30 days. It goes out of date when the thing it changes is edited in the meantime; "Approve" is then off.
  • When a step fails, the proposal shows what finished. On a proposal marked "Needs attention", "Retry" starts again at the step that stopped; a proposal marked "Failed" has no "Retry". When part of the change ran, "Roll back" makes a new proposal that undoes that part; that too needs approval.
  • After it ran, "Roll back" creates a new proposal to undo it. Some changes cannot be rolled back; the proposal page lists them.
  • "Simulate it first" is shown on a pending change to your AI configuration or to a knowledge item while Squid has simulations switched on, to the person who asked and to people who can approve it. It replays up to 40 recent first messages from customers on the proposed change and shows the result on the card. Nothing is sent to a customer.

Some cards ask you to tick a box before "Approve" turns on, for example when the change repeats wording a customer wrote or goes against a decision that Squid Agent has in memory.

Findings

"Findings" lists what Squid found in your support data, in six kinds:

  • "Automation opportunities": what to automate next. Squid scans for them once a week, on Monday after 04:00 UTC. Open opportunities are shown as a ranked table with the date of the last scan and, per row, "Tickets a month", "AI handles now", "Blocker" and "Handled without a person a month", an estimate with a range. A row from an earlier scan says "Not in the latest scan".
  • "Customer issues": a contact reason that rose for a product or a store. Squid checks once a week, on Monday after 05:00 UTC, comparing the latest week with the eight weeks before it.
  • "Knowledge": questions the AI could not answer, replies in which QA found invented facts, and a text that says something else than your return policy setting. Squid checks once a day.
  • "Investigations": conclusions you kept when Squid Agent explained a change, and the follow-up of a resolved incident.
  • "QA issues": failed QA checks that share a cause.
  • "Requests to Squid": what Squid cannot do yet for your stores, and Squid bugs your tickets ran into, each with Squid's status. Squid looks for new ones once a week. "Check now" looks at once; it works once per day (UTC) for the whole organization and needs 'View Reports'. A second click the same day answers "Squid already checked today", also when the first check failed.

Except under "Requests to Squid", the buttons "Open", "Dismissed" and "All" filter by state. The ordinary finding lists show 50 at a time with "Show more". The ranked opportunities table and "Requests to Squid" use separate lists without this button. Outside the ranked opportunities table, a line counts the findings you cannot see because of a store or a report, without naming them.

A finding's detail

Click a finding to open it under the list. It shows a summary, the numbers behind it, an "Evidence" label ("Observed correlation", "Temporal association", "Experiment-backed" or "Cause unknown"), the stores, the owner, when it was first and last seen, "Records behind it" with links, and its history.

  • An automation opportunity adds "Recommended path", "Blocker", "How it's estimated", and where Squid has them "Steps people take", "Why the AI hands over" and "Contacted again". The estimate is a number of conversations a month. Squid gives no hours or money.
  • A customer issue adds the contacts per week against what was sold, "What rose", "Suggested owner", "Affected orders" and "How it's found". It shows when contacts rose, not why.
  • A knowledge finding shows the AI's hand-over topic or, for a contradiction, the text and the setting side by side with a link to each.

The button at the bottom right starts a thread with a question typed in; you still press Enter. It says "Plan this with Squid Agent", "Draft the answer" or "Draft the fix". Three kinds of opportunity have no button: one about refunds or cancellations, which Squid does not recommend automating yet; one that needs an action or a connection Squid does not have yet, which belongs under "Requests to Squid"; and one where a policy has to be decided first.

Status, owner and dismiss

A finding moves through "Detected", "Reviewed", "Planned", "Implemented" and "Measured". A person can mark it reviewed, planned or implemented, and move it back from implemented to planned or from planned to reviewed. Squid moves it too: to "Planned" when a proposal is made for a detected or reviewed finding, unless that conversation read customer-written text, to "Implemented" when that proposal has run, back to "Planned" when that proposal is rolled back, and to "Measured" when the change's measurement is final. Nobody sets "Measured" by hand.

"Assign to me" makes you the owner. "Set owner outside Squid" records a team or a person by name, up to 200 characters. "Clear owner" removes the owner. "Dismiss" needs a reason of up to 500 characters, which everyone who sees the finding can read. A dismissed finding can be brought back with "Reopen"; it returns as "Reviewed".

You see a kind when you may see the report behind it: 'View Reports' for most kinds, 'View QA Reports' for QA issues. Changing a finding needs the permission for the area it is about: 'Edit AI Config' for automation opportunities and requests to Squid, 'Edit Variables' for knowledge, 'View Reports' for customer issues and investigations, and 'Edit QA Settings' for QA issues. Without it the detail says which permission is needed.

A finding is never a change. It leads to one only through a proposal.

Incidents

Squid checks your metrics and integrations every 15 minutes: contact volume, the share of conversations the AI hands over, refund rate, AI tool failures, CSAT and integrations whose login was rejected. Contact volume and AI handoffs are compared with the same weekday and hour over the last four weeks. Tool failures first use the same weekdays for that store and tool. Without enough history, Squid tries the organization's history for the tool, then its recent pooled failure rate, then a zero-rate comparison with a stricter threshold. CSAT counts low ratings, and both CSAT and refund rate use the last 24 hours against the corresponding periods in earlier weeks. Rejected integration logins need no historical comparison. The alert also needs a large enough rise and enough cases for that metric. When it is far outside its usual range for two checks in a row, Squid opens an incident with the status "Suspected" and notifies the people who manage incidents and can see its store. Signals on the same store join an active incident when its last signal was within two hours. A renewed signal of the same kind can also join an older active incident that had looked recovered.

The filter shows "Active", "Resolved", "Dismissed" or "All". A row shows the incident's number and severity, when it started, how many tickets are affected and its owner. A teammate whose access is limited to certain stores sees an incident only when all its stores are theirs; the list counts the others without naming them.

An incident's detail

  • A status bar: "Suspected", "Open", "Mitigating", "Monitoring", "Resolved".
  • The impact in numbers and a chart of the metric against its usual range.
  • "Likely causes", at most five, ranked, each with an evidence label. They can be a change made in the 24 hours before the incident started or after it started, a rise that is concentrated in one contact reason or channel, or a similar rise in several organizations at the same time. When nothing lines up, Squid says that nothing changed shortly before and that the cause is unknown. They are never stated as the cause. "Squid's explanation" links to a thread in which Squid Agent looked into it; Squid writes at most 10 of these a day per organization.
  • "Squid side" means the evidence points at Squid's platform and not at your setup. It is based on a change Squid made, or on a similar rise in several organizations; it does not prove that Squid changed something. The incident is then also listed under Findings → "Requests to Squid".
  • "Affected tickets": a sample, with "Open these … tickets in the inbox" and "Download CSV". The tickets are read-only here; tagging, assigning and closing happen in the Inbox. The list needs the 'View Conversations' permission, and the CSV also needs 'Export Conversations'. At most 5,000 tickets are linked to one incident.
  • "Response": who was notified, and the proposals made for this incident.
  • "Timeline": the first 200 recorded steps, oldest first, with who did them. Later steps are not shown once this limit is reached.

"Looks recovered — resolve?" appears after four checks counted as normal. A reading below the alert threshold, too little comparison data, or a missing reading can also count as normal. An incident that stays recovered for 24 hours is resolved by Squid.

What you can do

These need the 'Manage incidents' permission. Without it the detail says so.

  • "Confirm incident" moves a suspected incident to "Open". "Mitigating" and "Monitoring" set the status while you work on it.
  • "Dismiss" is for a suspected or open incident and needs a reason of up to 500 characters.
  • "Assign to me" makes you the owner.
  • "Notify people": tick up to 20 teammates and add a note of up to 300 characters. They get an in-app notification with a link. People who cannot see the store or who do not have 'Use Squid Agent' are not notified, and the page says how many.
  • "Run playbook", shown once the incident is confirmed, starts a thread with a question typed in, in which Squid Agent proposes who to notify and, where it helps, a temporary change to the AI. You still press Enter, and every change is a proposal.
  • "Resolve", then "Resolve and draft the postmortem": closes the incident. "What fixed it?" is optional. Squid writes a summary from the timeline, saves the follow-up as a finding under "Investigations", and suggests a memory of the incident that someone who manages memory approves.
  • "Reopen" brings a resolved or dismissed incident back.

Nothing on this tab writes to the affected tickets.

Goals

A goal is a number you want to reach by a date. The tab shows each goal with its target, deadline, guardrail, stores, owner and latest reading.

"Ask Squid to set a goal" starts a thread with "Set a goal:" typed in. There is no form on this page: you describe the goal and Squid Agent prepares a proposal. A goal has:

  • a name;
  • one metric: "Messages by AI", "AI handoff rate", "CSAT", "Average QA score", "First-response SLA compliance", "Resolution SLA compliance", "Average first response time" or "Average resolution time". When you ask for a metric Squid does not track goals on but knows the nearest one for, such as "AI resolution rate", it proposes that one and says so on the card;
  • a target, as "at least" or "at most" a value;
  • a deadline after today; for a new goal it must be within 730 days. Changing an existing goal checks that the date is after today, but does not apply that upper limit;
  • optionally a guardrail: a second metric that must stay above or below a bound;
  • the stores it covers, or all stores;
  • how many days each reading covers: 28 by default, from 7 to 90.

Anyone who uses Squid Agent can ask for a goal. Approving one, and approving a change to a goal or closing it, needs the 'Manage Squid Agent goals' permission. The person who approves a new goal becomes its owner. A goal is closed as "Achieved", "Missed" or "Dropped".

Squid reads each goal once a day, after 04:30 UTC, with the access of the goal's owner, and once right after the goal is set or changed. The reading is "Target met", "On track", "Off track" or "Unknown":

  • when the main number can be read, a breached guardrail makes the goal "Off track", also when the target is met;
  • "On track" means the reading is at or ahead of a straight line from the first reading to the target at the deadline. The first reading is "On track" only if the earlier checks did not already make it "Unknown", "Off track" or "Target met";
  • after the deadline a goal that did not meet its target is "Off track";
  • "Unknown" means there was no data, the number or the guardrail could not be read, or the owner left or lost access to the report or the stores. Squid then reads nothing with anyone else's access.

Every Monday each owner gets one notification that says how many of their goals are on track and off track. Someone who may not open the report behind a goal sees its status without the numbers. Goals about stores you cannot see are counted but not shown.

Under "Tasks" are follow-up tasks, filtered by "Open", "Done", "Closed" or "All". A task is created by an approved proposal or by an incident's postmortem. Each task has a check, shown as "Done when: …", such as a rule that is live, a setting with a certain value, a published FAQ entry, a proposal that ran or a finding that was implemented. Squid looks at every open task once an hour and marks it done when its check is met. Nothing said in a thread closes a task. An overdue task shows "Overdue", and its owner gets a reminder at most once a day.

"Close…" closes a task by hand. It needs a note of at least 3 and at most 500 characters, and it is for the task's owner or someone with 'Manage Squid Agent goals'. The task then shows "Closed by a person" with the note, and the close is recorded in the audit log.

Memory

Memory is what Squid Agent remembers about how your team works, so you do not have to repeat it. Memories are notes, not instructions: they never change anyone's access, and Squid Agent still looks up every number and setting when it answers.

Memories have three reaches, each with its own tab: "Only me" (used only in your own threads), "Stores" and "Organization". You add one by telling Squid Agent in a thread, for example "remember that…" or "from now on…". What happens then depends on what it is and who asks:

  • Something for yourself is saved at once when you asked for it in your own words, and shows "Remembered for you" with "Undo". Something Squid Agent inferred becomes a "Remember this?" card first.
  • A fact or context for a store or the organization (a business fact, a person and their role, a term, an incident, a goal, a pattern, an open question) is saved at once when you hold 'Manage Squid memory and shared skills'. From anyone else it becomes a suggestion that someone with that permission approves. A memory for the whole organization from someone whose access is limited to certain stores is always a suggestion, and you cannot ask Squid Agent to remember something for a store you have no access to.
  • A decision, a constraint or a preference for a store or the organization changes how Squid Agent behaves, so it is always a suggestion that someone with 'Manage Squid memory and shared skills' approves. One exception: a temporary constraint with an end date that such a person asked for is saved at once.
  • Wording that repeats what a customer wrote is never saved. In a conversation that read customer-written text nothing is saved at once; everything becomes a card.
  • A memory that contradicts an existing one, or that reads like an instruction, is never saved at once. It becomes a card.

"Undo" on a memory that was saved at once works for 30 minutes for the person who asked. After that, forget it from this page. An organization keeps at most 1,000 active memories.

Changing a memory

On a memory you can:

  • "Pin" it to give it priority in the memories loaded for an answer ("Always remembered"). Store memories are loaded for the stores in the question or page, up to three stores; pinning does not load every store into every answer. Each reach has room for a limited amount of pinned text; the meter at the top shows how much is used. When it is full, unpin something first.
  • Edit it. The form has "Text" (3 to 500 characters), "Type" and "Until (optional)", an end date after which it is no longer current. Squid Agent can still recall it as expired history. The types are "Business fact", "Decision", "Preference", "Person and role", "Glossary", "Constraint", "Incident", "Pattern", "Goal" and "Open question". Every edit is saved as a new version; earlier versions stay in the memory's history.
  • Forget it. Squid stops using it, and it can be restored from "History".

Your own memories are yours to change, and nobody else can read them, the Owner included. Changing store and organization memories needs 'Manage Squid memory and shared skills' and access to the stores the memory is for. A person with that permission whose access is limited to certain stores cannot change organization memories.

Open a memory to see where it came from ("Sources", with a link to the thread) and its history. A memory that has passed its review date is marked "Still true? It is due for a review."; one marked "Expired" has passed its end date. While "Memory" is on, Squid checks this every night, and preferences and patterns that were not used for a year and are not pinned are forgotten automatically. No other type is forgotten automatically.

Pending, Conflicts and History

"Pending" holds suggestions waiting for approval. On a card for someone else's suggestion, "Approve" saves it, "Edit" lets you change the text, type or end date and then "Save and approve", and "Reject" turns it down. On your own suggestion the buttons are "Save" and "Not now"; when you may not approve it yourself they are "Send for approval" and "Not now". This tab is the one place with "Approve selected", for up to 50 suggestions at once; a suggestion that read customer text has to be opened and approved on its own. Suggestions expire after 30 days.

"Conflicts" shows suggestions that contradict an existing memory, side by side: "Replace" (the old one stays in History), "Both are true" or "Keep old".

"History" lists changes and forgotten memories, with "Restore". Each list shows at most 200 rows.

Switches, export and erase

Two switches, for people with 'Manage Squid memory and shared skills':

  • "Memory": off, Squid Agent neither saves nor uses memories during answers. Existing ones are kept and can still be edited from this page.
  • "Learning (suggestions and nightly review)": on, Squid reviews the day's finished threads each night and suggests memories for stores and the organization. It reads only what people asked and what Squid Agent answered, never customer text. It never saves anything itself; the suggestions wait in "Pending" for anyone who may approve them.

While "Memory" is on, people who manage memory get a digest every Monday, unless they switched that notification off.

"Export my memories" downloads your personal memories. "Export organization memory" downloads every organization and store memory with its history, without anyone's personal memories; it is for the Owner. "Erase all" is for the Owner too and asks you to type ERASE ALL MEMORY before "Erase all memory" turns on. It erases every memory, personal ones included, and cannot be undone.

Skills

A skill is a saved, named list of look-ups that Squid Agent can run again, such as a weekly review. A skill has 1 to 8 steps, and every step only reads; a skill never changes anything.

  • "Your skills" are yours alone. Nobody else sees or runs them, the Owner included. You make one with "Save as skill" under an answer: Squid keeps up to eight of the look-ups that answer made that succeeded and can be repeated; other steps are left out. Recent dates are stored relative to the day you asked, so "last 7 days" stays the last 7 days. A date more than 400 days before the question, or after it, stays fixed.
  • "Organization skills" are shared with everyone who uses Squid Agent. They run with the access of the person who runs them, so two people can get different results. Someone who cannot reach a store a skill reads sees its name and "This skill reads a store you don't have access to", not its steps.
  • "Browse templates" shows ready-made skills from Squid. "Install" copies one to your skills. When Squid publishes a newer version, the skill shows "Update available" and "Review update" shows what changed. Your own changes are kept unless you pick Squid's version for a step. For your own skill "Apply the update" works at once; for an organization skill it becomes a proposal.

"Save this as a skill?" can appear after the same repeatable look-ups succeeded twice for you within 30 days. Runs rated "Not helpful" and runs that prepared a change do not count. It is not offered when you already have a skill with those same look-ups. "Not now" stops it asking about that look-up for 30 days.

Open a skill to see its steps and version history. "Run now" opens a new thread with the skill's question typed in; you still press Enter, and the run uses one answer. "Schedule", on a skill of your own, runs it on a timetable. "Delete" removes the skill. Its schedules pause when they next check and find it missing. You can delete your own skill even after sharing it; deleting someone else's organization skill needs 'Manage Squid memory and shared skills'.

Saving a skill from an answer is immediate, as version 1. When a look-up in that answer was chosen after Squid Agent had read customer-written text, the save becomes a proposal that you approve after checking the wording. Restoring an earlier version of your own skill happens at once, adds a new version and can be undone. "Share with organization" happens at once, with "Undo", when you hold 'Manage Squid memory and shared skills' and may read everything the skill reads; from anyone else it becomes a proposal that someone with that permission approves. Undoing your own immediate share works for 24 hours. It is refused if another person has scheduled or changed the skill, or if you already have a personal skill with that name. Restoring a version of an organization skill is always a proposal that someone with that permission approves.

Schedules

A schedule runs something for you at set times. There are three kinds:

  • "Briefing": a question Squid Agent answers on a timetable. "New schedule" asks for the question (up to 2,000 characters), "Repeat" and "Timezone". It can read the stores you have access to, according to the question.
  • "Skill": a skill run on a timetable, made with "Schedule" on a skill. It always runs the latest approved version.
  • "Alert": a check on one metric. "New alert" asks for the metric ("First-response SLA breaches", "Resolution SLA breaches", "New tickets", "AI handoffs to a person" or "Refunds"), one store or "All your stores", the timezone, and "More than" or "Less than" a value. The window is always "Per day". Squid checks the number every 15 minutes without the AI. The timezone chooses the calendar date and the once-a-day reset. It does not shift the report's day boundaries into that timezone: the check passes that date to the report, with UTC bounds where the report accepts times. When the condition holds, Squid Agent writes an answer that explains it, at most once per day in the timezone you chose.

"Repeat" offers "Weekdays at 08:00", "Every Monday at 08:00", "Every day at 07:30" and "First day of the month at 09:00". Ask Squid Agent in a thread for another time: a timetable can be every day, weekdays, one day of the week or one day of the month (the 1st to the 28th), at a time in your timezone.

Creating or changing a schedule makes a card that only you can approve, because the schedule runs as you. Your access is checked again before every run. If you lost a permission, a store or the skill it needs, the schedule is paused and says why. "Pause", "Resume" and "Delete" work at once; "Resume" checks your access first.

Results arrive in the app only: a thread, and a notification unless you muted that kind. A briefing, a skill run and an alert that fires each use one of your organization's daily answers; an alert check that does not fire uses none. A run is skipped, not made up later, when the daily limit is reached, when Squid Agent is paused, or when the organization reaches its scheduled-run limit for that UTC day, 50 by default; Squid can change this limit. Runs are also skipped when Squid Agent is switched off or the organization has no active plan. A missed time is not caught up more than once.

You only see your own schedules. Click one to see its runs and open their threads.

Activity

"Activity" lists the questions that were asked on one day: when, who, the question, the tools Squid Agent used, how long it took, the result and the feedback given. The day is counted in UTC and you can pick an earlier day.

With the 'View Everyone's Squid Agent Threads' permission you see every run in the organization; otherwise only your own. A run that used a store or a report you may not see shows who asked and when, without the question. The result is "Succeeded", "Failed", "Cancelled", "Stopped by Squid" or "No metric" (the question requested an unavailable metric and the successful answer has no supporting numbers), and a run that is still going shows "Running". Runs started by a schedule or by the nightly review are marked "Scheduled" and "Nightly". The list never shows what an answer cost.