Contents Mastering AI in Your Business

    Agent Build Guide · Updated

    Mastering AI in Your Business

    You start automating your business by picking one task you already do the same way every time and building one agent for just that task. Not the whole business, one task, and one agent made of a few plain parts: something that starts it, a couple of connections, and a written job description. This guide covers what to automate first, what an agent is made of, when to build versus when to just write it down, the first agents to build, and what to guard against.

    The platform used throughout is Gumloop, a tool for building and running AI agents that connect to the apps you already work in. It runs on a 14-day trial, then a paid plan.

    Start the two-minute assessment

    01 · What to automate first

    What should I automate first?

    Automate the work you already do the same way every time. Rule-based, repeatable, low-ambiguity tasks make the best first builds, because you can tell inside a single run whether the agent got it right or not. Anything that turns on a judgment call, or on a relationship where the wrong tone costs you something real, is a poor first candidate.

    Three questions, about two minutes. Every answer routes to a specific build.

    What eats the most repetitive time in your week right now?
    Is this something just you would use, or would your team use it too?

    If you would rather just read the routing:

    What eats your weekStart hereWhy
    Replying to the same kinds of messagesInbox follow-upIt drafts the reply. You still hit send.
    Rewriting one piece of content for every channelContent repurposingIt does the reformatting. You keep the words.
    Looking people or companies up before a conversationProspect researchIt hands you the brief before you go looking for it.
    Moving the same data between tools by handAdmin data-entryThis is the job agents are best at. Start here.
    Reading to stay currentWeekly research digestIt takes the first pass so you only read what matters.
    None of these, mostly judgment and conversationWeekly research digestIt only reads and summarizes, so it is the lowest-risk agent to try. If scheduling or data entry is more your edge than reading, start with admin data-entry instead.

    02 · The parts

    What is a Gumloop agent actually made of?

    A Gumloop agent is an assembly of six kinds of parts, a Trigger, Connectors, Skills, Knowledge Sources, Subagents, and Abilities, rather than one large model that somehow already knows your business. Everything the agent knows about how you work, it knows because one of those six parts put it there, and you can open any of them and see exactly what it says.

    1

    Trigger

    The condition that starts the agent running: a schedule, an event in a connected app, or a check you describe in plain language. Gumloop offers three kinds: Create With AI for custom polling Triggers you describe yourself, Event-Based Triggers built for a single service, and Scheduled Triggers that run on a set schedule or once at a set time.

    2

    Connector

    An authorized link to an outside tool the agent reads from or writes to. Gumloop documents more than 150 integrations, including Gmail, Outlook, Google Calendar, Sheets, Docs and Drive, Slack, Notion, Airtable, HubSpot, Salesforce, LinkedIn, Instagram, YouTube, Canva, Gamma, Beehiiv, Stripe, QuickBooks, Cal.com, Luma, Zoom, Google Analytics, and Google Search Console.

    3

    Skill

    A reusable set of instructions, and optionally templates and scripts, that teaches an agent how to do one specific task. You can build one three ways: Create With AI, Write Skill Instructions, or Upload Files.

    4

    Knowledge Source

    A body of your own material the agent searches before it answers. Gumloop's Brain connects Notion, Google Drive, Slack, GitHub, Confluence, Zendesk, and Salesforce, plus uploaded files. Attach knowledge and the agent gains a Search Company Brain tool that returns relevant snippets with citations instead of guessing.

    5

    Subagent

    A separate agent your agent calls to handle a distinct job with its own tools. An agent can clone itself, which appears in selection lists as (Me), with a depth limit of one so clones cannot spawn clones. It can also invoke a different named agent, with no depth restriction.

    6

    Ability

    A built-in capability the agent has without connecting anything: Web Search, Web Fetch, Image Generation, a Code Sandbox that runs Python and shell commands, and Ask Question, which pauses and asks you for input through a structured choice card.

    Walking the builder screen

    Agent Preferences is where a build starts. Pick the model that powers the agent, then use the Add instructions for the agent box to write what it is, who it works for, and how it should behave. This box is the agent's job description, and most of what people expect from a Skill actually belongs here.

    Beneath the instructions sits the Allow Self-Updates toggle, which lets the agent revise its own instructions. It is on by default. Leave it on and you can correct the agent in conversation, tell it to update its instructions so it always works that way, and watch the change land in the box you just wrote.

    Triggers is where you decide what starts it. The panel carries an AI Managed setting, and the agent can create and manage its own Triggers during a conversation, so scheduling can be a thing you ask for rather than a form you fill in.

    Connectors is the authorization step, shown with an AI Discovery setting. Add only what this agent needs. An agent holding every Connector in the company is the signal that you should have split it into specialists.

    Skills holds the reusable procedures, with an AI Skill Editing setting on the panel. Skills can be updated by the agent itself when you correct it, or edited by hand by anyone with agent editing access.

    Knowledge Sources is where you attach the documents and workspaces the agent should read before it answers. Subagents is where you add the other agents it can call, including its own clone as (Me). Abilities holds the built-in switches, Web Search and Web Fetch among them.

    Now see how the parts combine

    03 · Build it or write it

    When do I create a Skill, use a Subagent, or just write it into the instructions?

    Write the logic into the instructions when it governs everything this one agent does and fits in a few lines. Make it a Skill, a reusable set of instructions that teaches an agent one specific task, when it is a standing procedure you would otherwise paste in as a long document, or when a second agent will need the same procedure. Hand it to a Subagent, a separate agent called to do a distinct job with its own tools, when the work is a different job with its own steps, its own connections, and its own back-and-forth.

    Gumloop's documentation states the same test more bluntly: "If you are about to paste a long SOP into the agent instructions, make a skill. If you are about to give one agent every connector in the company, split specialists and call them as subagents."

    ObjectBuild it whenWhat it isWhere it runs
    InstructionsThe rule applies always, to this agent only, and fits in a few linesThe text in the Add instructions for the agent boxEvery turn, automatically
    SkillThe procedure is long, repeatable, or shared across agentsA reusable set of instructions, plus optional templates and scriptsInside the current conversation, when relevant
    SubagentThe work is a separate job with its own tools and its own reasoningA different agent, or a clone of this one shown as (Me)In its own conversation, then reports back
    App RulesAn action must be impossible, not merely discouragedAn enforcement policy, set by agent editors or org adminsAt the tool call, before it executes

    App Rules are the fourth object, and the one people skip. Skills suggest, App Rules enforce. Gumloop's own comparison page puts it this way: "A skill can ask the agent not to delete Slack messages. An app rule can stop the delete call." Anything you would be upset to discover after the fact belongs in an App Rule, not in a politely worded Skill.

    Worked example: inbox follow-up, instructions

    The inbox follow-up agent drafts replies to the messages you answer the same way every week. "Draft, never send. Match the voice in these five examples. Flag anything about pricing for me instead of answering it." That is three lines, it applies to every message the agent touches, and it belongs in the instructions box. Building a Skill for it would add a file to maintain and change nothing about the output.

    Worked example: content repurposing, a Skill

    Content repurposing goes the other way. Every format you publish carries real rules about length, structure, and how a caption differs from a newsletter section, and writing all of that into the instructions box means pasting a document into a field meant for a job description. That is a procedure, and you will want the same procedure the day you build a second agent for a different channel. Make it a Skill.

    Worked example: prospect research, a Subagent

    Prospect research does not belong on the inbox agent. Research needs the web, your CRM, your calendar, and your notes, and bolting all of it on leaves you with one overloaded agent holding every Connector you own. Keep research as its own agent with its own connections. The inbox agent calls it as a Subagent when a reply needs context, and it hands back the brief.

    The admin data-entry agent is where App Rules earn their keep. It writes to live systems, so the rule that stops a delete or an overwrite has to be enforcement, not instruction.

    Jump straight to a build: inbox follow-up, content repurposing, prospect research, admin data-entry, or the weekly research digest. Not sure which fits your week? Take the assessment.

    See the five builds

    04 · The builds

    Five agents to build first

    These five are common, bounded, low-risk jobs, and each one sits in a different function of running a business: your inbox, your content, your prospect research, your back office, and your industry reading. None of them lets the agent send, publish, or spend anything on its own.

    Each recipe uses the same ten fields so you can compare them before you pick one. The instructions block in each is written to paste straight into the "Add instructions for the agent" box in Agent Preferences, with your own details swapped in. Each recipe also names the Skills to build and why; the test is the one from the decision table: a reusable process becomes a Skill, a one-off detail stays in the instructions. New to the parts vocabulary? The anatomy section defines all six in a minute of reading. Build one, watch it run, then build the next from what you learned.

    01/5

    Inbox follow-up

    This agent reads incoming email, recognizes the few message types you answer the same way every time, and writes a draft reply you review and send yourself. It is for anyone whose inbox holds a steady flow of routine requests: meeting asks, invitations, pricing questions, and requests for a document you have already sent to someone else.

    Trigger Event-Based, on a new email in Gmail or Outlook
    Connectors Gmail or Outlook, Google Calendar, Google Drive
    Expected build time 60 to 90 min

    Connectors, and what each one does here

    • Gmail or Outlook. Reads the incoming message and saves its reply as a draft on the same thread. Gmail's documented capability includes managing drafts, which is the basis of this recipe.
    • Google Calendar. Reads your availability so a meeting reply proposes times you can keep.
    • Google Drive. Locates the document a sender is asking for instead of describing one from memory.

    Skills to create, and why

    A Skill is a reusable instruction pack an agent follows for a specific task. Build two. The first is a Reply Voice Skill through the Upload Files path: upload five emails you have already sent, covering your greeting, your sign-off, your paragraph length, and how you say no. Uploading your own sent mail is the difference between a reply that sounds like you and one that sounds like a template. The second is a Routine Request Types Skill through Write Skill Instructions, listing the four or five categories this agent may draft for and the standard shape of each answer.

    Setup steps in order

    1. Create a new agent and name it for the job, not for the tool.
    2. In Agent Preferences, choose your model and paste the instructions block below into "Add instructions for the agent."
    3. Add the Gmail or Outlook Connector, then Google Calendar and Google Drive.
    4. Create the Reply Voice Skill by uploading five of your own sent emails.
    5. Create the Routine Request Types Skill by writing out your categories.
    6. Run the agent manually against three real messages already in your inbox and read every draft.
    7. Set the Event-Based Trigger only after those three drafts pass.

    Paste this into "Add instructions for the agent"

    Paste this
    You draft replies to routine email for me. You never send.
    
    When a new email arrives, read it and decide whether it is one of the kinds I handle the same way every time: a meeting request, a speaking or podcast invitation, a pricing question, or a request for a document I have sent before. If it is not one of those, do nothing and leave the message alone.
    
    For a routine email, write a reply and save it as a draft on that thread. Follow the Reply Voice Skill for greeting, sign-off, and paragraph length. Answer the question that was asked in the first sentence, then add context only if the answer needs it. Keep the draft under 150 words unless the sender asked something that needs more. Do not use exclamation points. Do not use em dashes.
    
    Never send. Never delete, archive, or label anything. If a message mentions a refund, a contract, a legal issue, a complaint, or money owed, stop and use Ask Question to show me the message and ask how I want to handle it. If you cannot tell what the sender wants, use Ask Question instead of guessing.
    What to check before you turn it on
    • Run it manually on three real messages and read every draft word for word.
    • Confirm the reply landed as a draft on the thread and nothing left your outbox.
    • Send it a message mentioning a refund and confirm it stops and uses Ask Question.
    • Check one drafted meeting reply against your calendar for that day.
    Time math (illustrative, swap in your own counts)

    If you answer 20 similar emails a week at 6 minutes each, that is 120 minutes. If reviewing and sending a draft takes 2 minutes, 20 drafts cost you 40 minutes. The agent gives back about 80 minutes a week in that example. Count your real messages for one week before deciding what this is worth.

    You're done when

    02/5

    Content repurposing

    This agent takes one finished piece of your writing and produces channel-ready drafts of it, without adding a claim you did not write. It is for anyone who writes something once and then rewrites it four more times for a newsletter, a post, and short social copy.

    Trigger Create With AI, watching a named folder for a new document
    Connectors Google Drive, Google Docs (Beehiiv later, behind your yes)
    Expected build time 45 to 75 min

    Connectors, and what each one does here

    • Google Drive. Watches the folder where you drop a finished draft.
    • Google Docs. Reads the source and writes all output into one new document you review.
    • Beehiiv. A verified Connector for newsletters, documented for publishing posts and managing subscribers. Publishing is exactly what you do not hand over on a first build, so add this Connector only when you are ready and keep the publish step behind your own yes.

    Canva and Gamma are verified Connectors too, if a repurposed piece needs to become a design or a deck. Leave both out of your first build.

    Skills to create, and why

    Build a House Format Skill through Write Skill Instructions, specifying the exact length and shape of each output: newsletter length, post length, how many short social variants, how many subject-line options. Build a second Source Discipline Skill through Create With AI, stating that every claim in an output must already exist in the source. Gumloop supports a faster path for both: do the task once in chat, correct it until it is right, then tell the agent to turn that into a Skill.

    Setup steps in order

    1. Create the agent and paste the instructions block below into "Add instructions for the agent."
    2. Add the Google Drive and Google Docs Connectors.
    3. Create a folder for this purpose and put one finished piece in it.
    4. Create the House Format Skill with your real lengths and formats.
    5. Create the Source Discipline Skill.
    6. Run the agent manually on that piece and read every output against the source.
    7. Set the Create With AI Trigger to watch the folder once the manual run is clean.

    Paste this into "Add instructions for the agent"

    Paste this
    You turn one finished piece of my writing into channel-ready drafts. You never publish.
    
    When a new document appears in the Repurpose folder, read the whole thing before you write anything. Produce all output in one new Google Doc titled with the source title plus "repurpose drafts." Include four sections: a newsletter version, a long-form post, three short social posts, and five subject-line options. Follow the House Format Skill for the exact length of each.
    
    Use only the claims, numbers, examples, and stories that already appear in the source document. Do not add a statistic, a client story, or a quotation that is not in the source. If a claim in the source carries no attribution, leave that claim out of the shorter formats rather than repeating it stripped of its context. Do not use em dashes.
    
    Do not publish to Beehiiv. Do not post anywhere. Do not generate images unless I ask. When the drafts are ready, give me the document link and stop. If the source document is unfinished or you cannot tell what it argues, use Ask Question rather than writing around the gap.
    What to check before you turn it on
    • Run it on one piece you know well and trace every claim in every output back to the source.
    • Confirm it wrote to a new document and changed nothing in your original.
    • Confirm nothing was published or posted anywhere.
    • Read the short social drafts specifically, since compression is where invented claims appear.
    Time math (illustrative, swap in your own counts)

    If you publish 4 pieces a month and reformatting each into four channels takes 50 minutes, that is 200 minutes. If editing four drafted versions takes 15 minutes per piece, that is 60 minutes. The agent gives back about 140 minutes a month in that example. Count your own last month before trusting that shape.

    You're done when

    03/5

    Prospect research

    This agent writes you a one-page brief on every external person you are meeting tomorrow, using your own records first and public sources second. It is for anyone who does the research at the last minute, or not at all.

    Trigger Scheduled, every weekday morning ahead of your first meeting
    Connectors Google Calendar, Google Docs, HubSpot or Salesforce
    Expected build time 60 to 90 min

    Connectors, and what each one does here

    • Google Calendar. Reads tomorrow's meetings and pulls attendee names and email domains.
    • Google Docs. Writes one brief per meeting into a document you open before the call.
    • HubSpot or Salesforce. Reads the existing record so the brief includes what your team already knows.

    Abilities and Knowledge Sources

    An Ability is a built-in capability an agent uses without a Connector. This recipe needs two: Web Search to find current public information, and Web Fetch to read a page you named. It also needs a Knowledge Source, which is your own material attached to the agent so it can search and cite it. Attach Google Drive, or Notion if your notes live there. Both are verified Knowledge Sources, and an agent with knowledge attached gets a Search Company Brain tool that returns answers with citations instead of guessing. Point it at your own records first and the open web second, in that order.

    Skills to create, and why

    Build a Meeting Brief Format Skill through Write Skill Instructions, fixing the five sections every brief contains and the rule that each factual line carries a source link. A brief with no fixed shape is one you have to read twice. Build a second Verification Rule Skill, same path, stating that anything it cannot confirm gets written as "not verified" rather than filled in with something plausible.

    Setup steps in order

    1. Create the agent and paste the instructions block below into "Add instructions for the agent."
    2. Add the Google Calendar and Google Docs Connectors, then your CRM Connector.
    3. Turn on the Web Search and Web Fetch Abilities.
    4. Attach your Google Drive or Notion as a Knowledge Source.
    5. Create the Meeting Brief Format and Verification Rule Skills.
    6. Run it manually against tomorrow's calendar and check every source link.
    7. Set the Scheduled Trigger for early morning once two manual runs come back clean.

    Paste this into "Add instructions for the agent"

    Paste this
    You write me a one-page brief on each person I am meeting tomorrow. You research only. You never contact anyone.
    
    Read tomorrow's calendar. For every meeting with an external attendee, collect the attendee names and email domains. Search Company Brain first for anything we already have: prior notes, proposals, past threads, existing CRM records. Only after that, use Web Search and Web Fetch to check the organization's own website and the attendee's public professional profile.
    
    Write one Google Doc per meeting, following the Meeting Brief Format Skill: who they are and their current role, what the organization does in two sentences, what has changed there recently, what we have already discussed with them according to our own records, and three questions I could ask.
    
    Put a source link on every factual claim. If you cannot verify something, write "not verified" and move on. Do not fill a gap with a plausible guess. Do not use em dashes.
    
    Never email, message, connect with, or follow anyone. Never write into the CRM record; you read from it only. If a meeting has no external attendee, skip it without comment.
    What to check before you turn it on
    • Run it on tomorrow's calendar and click every source link in the brief.
    • Pick one person you already know well and check whether the brief got them right.
    • Confirm the CRM record is unchanged after the run.
    • Confirm it wrote "not verified" somewhere. Five confident pages with no gaps is the sign it is filling them in.
    Time math (illustrative, swap in your own counts)

    If you have 6 external meetings a week and researching each properly takes 15 minutes, that is 90 minutes. If reading a prepared brief and adding your own notes takes 4 minutes, 6 briefs cost you 24 minutes. The agent gives back about 66 minutes a week in that example, and the larger gain is the meetings you would have walked into unprepared. Use your own meeting count.

    You're done when

    04/5

    Admin data-entry

    This agent takes new intake rows from a spreadsheet, standardizes them, and files them into your database without creating duplicates. It is for the person still copying form responses into a second system by hand.

    Trigger Event-Based on a new row in Google Sheets, or Create With AI polling the sheet hourly
    Connectors Google Sheets, Airtable, Slack
    Expected build time 45 to 90 min

    Connectors, and what each one does here

    • Google Sheets. Reads new intake rows and marks each one processed.
    • Airtable. Creates or updates the record, and never deletes one.
    • Slack. Posts a short summary at the end of each batch so the work stays visible to you.

    Skills to create, and why

    Build an Intake Mapping Skill through Write Skill Instructions, and be specific: which sheet column maps to which Airtable field, the exact allowed values for every category field, the date format, and how names and organizations get capitalized. This is the highest-value Skill in the whole set, because data entry breaks on details rather than on judgment. Build a second Duplicate Rule Skill defining what counts as the same record, usually a matching email address, and what the agent does when two records look like the same person under different addresses.

    Setup steps in order

    1. Create the agent and paste the instructions block below into "Add instructions for the agent."
    2. Add the Google Sheets and Airtable Connectors, then Slack.
    3. Write the Intake Mapping Skill with your real column-to-field map and allowed values.
    4. Write the Duplicate Rule Skill.
    5. Duplicate your Airtable base, or point the agent at a test table, so a first run cannot damage live records.
    6. Run it manually on five real rows, including one incomplete row and one deliberate duplicate.
    7. Point it at the live table and set the Trigger once those five rows come back correct.

    Paste this into "Add instructions for the agent"

    Paste this
    You move new intake rows into our Airtable records and flag anything you are unsure about.
    
    When a new row appears in the intake sheet, read every field. Standardize it using the Intake Mapping Skill: full name in one field, organization name spelled the way the organization spells it on its own site, email lowercased, request type set to one of the allowed values, date as YYYY-MM-DD.
    
    Before you create anything, search Airtable for an existing record with the same email address, following the Duplicate Rule Skill. If one exists, update that record and note in your summary that this was an update. Never create a second record for the same person.
    
    If a required field is empty, if the request type does not clearly match one of the allowed values, or if two records look like the same person under different email addresses, do not guess. Use Ask Question and show me the row alongside your two best options.
    
    Never delete an Airtable record. Never change a field you were not asked to fill. Do not use em dashes. At the end of each batch, post one Slack message: rows added, rows updated, rows waiting on me.
    What to check before you turn it on
    • Run it against a copy of your table, never the live one, on the first pass.
    • Feed it a row with a missing required field and confirm it stops and uses Ask Question.
    • Feed it a row for someone already in the database and confirm it updates rather than duplicates.
    • Open five records it created and check every field against the source row.
    Time math (illustrative, swap in your own counts)

    If 40 intake rows a week each take 3 minutes to enter and check, that is 120 minutes. If the agent files most of them cleanly and flags the rest for you at 2 minutes each, your part is the flagged ones plus the Slack summary. Count your real rows for one week, and count the flagged ones too, because the flag rate decides whether this is worth building.

    You're done when

    05/5

    Weekly research digest

    This agent reads your field for you once a week and writes one briefing on what changed. It is for anyone who keeps meaning to stay current and keeps running out of week, and it is the lowest-risk agent on this page because it only reads and writes to a document you own.

    Trigger Scheduled, weekly, Monday at 7am
    Connectors Google Docs, Google Drive, Slack
    Expected build time 45 to 60 min

    Connectors, and what each one does here

    • Google Docs. Writes each week's digest into a new dated document.
    • Google Drive. Files the digest in one folder so the archive builds itself.
    • Slack. Posts the link, which is the difference between a document that gets read and one that quietly accumulates.

    Abilities

    This recipe runs on Web Search and Web Fetch, built-in capabilities the agent uses without a Connector. Web Search finds what published this week; Web Fetch reads a specific page you named.

    Skills to create, and why

    Build a Digest Sources Skill through Write Skill Instructions, naming the publications, sites, and topics that count for your field and the ones that do not. Without it, the agent decides what your field is, and it will decide wrong. Build a second Digest Format Skill fixing the shape of the briefing: three sentences on the most important development, then up to eight items, each with a headline, two sentences, a publication date, and a link.

    Setup steps in order

    1. Create the agent and paste the instructions block below into "Add instructions for the agent."
    2. Add the Google Docs and Google Drive Connectors, then Slack.
    3. Turn on the Web Search and Web Fetch Abilities.
    4. Write the Digest Sources Skill with your real source list.
    5. Write the Digest Format Skill.
    6. Run it manually and check every link and every publication date.
    7. Set the Scheduled Trigger for Monday morning once one manual run comes back clean.

    Paste this into "Add instructions for the agent"

    Paste this
    You write me one weekly briefing on what changed in my field. You read and summarize only.
    
    Every Monday morning, use Web Search and Web Fetch to check the sources named in the Digest Sources Skill. Consider only material published in the last seven days. Skip anything older. Skip anything you covered in last week's digest.
    
    Write one Google Doc titled with the date, following the Digest Format Skill. Open with three sentences naming the single most important development of the week and why it matters to a small business owner in my field. Then list up to eight items. Give each item a headline, two sentences of summary, the publication date, and a link.
    
    Attribute every claim to the source that made it. If two sources disagree, say so rather than picking a side. If the week was quiet, say the week was quiet and list fewer items. Do not pad the list to reach eight. Do not include an item whose date you could not confirm. Do not use em dashes.
    
    Do not email this to anyone. Do not post it publicly. Do not publish it to any newsletter. Post the document link in Slack and stop.
    What to check before you turn it on
    • Run it manually and open every link in the output.
    • Check every publication date, because a stale item dressed as news is this recipe's main failure.
    • Confirm nothing was emailed, posted, or published.
    • Read the three-sentence opening and ask whether it names something that happened this week.
    Time math (illustrative, swap in your own counts)

    If you spend 45 minutes a week scanning sources and you skim a prepared digest in 10 minutes, the agent gives back about 35 minutes a week in that example. Use your own honest count.

    You're done when

    09 · Operating rules

    What should the agent never be allowed to do on its own?

    An agent needs a human yes before any action it cannot take back: sending anything to a person outside your company, spending money, deleting data, or changing who has access to what. Everything short of that line can run unattended once you have watched it work a few times. The line matters more than the list, because an agent that drafts a bad email costs you two minutes of editing and an agent that sends one costs you a relationship.

    Write the rules before the agent runs. The Add instructions for the agent box is blank at the start of a build, which is the cheapest moment you will ever have to decide what this agent is not allowed to do.

    Four rules cover most of what a small team needs, and each one has a place in the product where the enforcement lives.

    Rule 1. Draft, never send.

    Write it into the agent's instructions in plain words: produce a draft, put it where I will see it, stop there. Then make the output step match, routing the result to a draft or a message to you rather than to a send action. An instruction is a default, not a lock, which is the whole reason rule 3 exists.

    Rule 2. When the rule runs out, ask.

    Turn on the Ask Question Ability, one of the agent's built-in Abilities (an Ability is something the agent can do on its own, separate from the outside tools you connect to it). It pauses the run and hands you a structured choice instead of letting the agent guess. Use it at every fork you did not write a rule for, such as which client an ambiguous invoice belongs to.

    Rule 3. Hard-block what must never happen.

    App Rules are enforcement policies rather than guidance, and they stop the call the agent tries to make. Deleting records, touching a billing tool, posting anywhere public: if you would not want to explain it to a client, it belongs in an App Rule and not in a paragraph of instructions.

    Rule 4. Approve tools one at a time.

    Gumloop documents four approval mechanisms: App-Level Settings, Per-Tool Custom Controls, App Rules, and the Ask Question Ability. Rule 4 is the first two. Connect only the tools this agent's job requires, then set approval on the individual tools that write, send, or spend, rather than approving an entire app in one click.

    The difference between a Skill and an App Rule is the one that trips teams up. A Skill is a reusable set of instructions that teaches an agent how to do a specific task, and what it produces is a suggestion the agent can override. An App Rule cannot be overridden. Teams write a careful Skill, feel covered, and have installed a preference rather than a control.

    Rolling this out past yourself changes the problem from what the agent may do to what your team will keep doing.

    Start with team rollout

    10 · Past yourself

    How do I roll this out to a team without it dying in month two?

    Rollouts die because the agent had no owner once the demo ended, and the fix is one name attached to every agent before anyone else touches it, plus a standing check-in that survives a busy week. Five moves hold a small team past month two.

    1. Learn it yourself first.

    Build one agent for one of your own recurring tasks and run it for one to two weeks before you introduce it to anyone. Write down each decision while you are making it, including the ones you got wrong, because that document becomes your training material. Time the task by hand before you build and time it again after. If the manual version takes 40 minutes and the agent version takes 8, the saving is 32 minutes per run, and at two runs a week that is 64 minutes back. Those are your numbers from your own work, which is a different conversation than a vendor's slide.

    2. Hold a weekly AI standup, 15 to 20 minutes, same slot every week.

    Keep the agenda tight: five minutes on what worked, stated in minutes saved rather than impressions; five minutes on what failed and why it was cut; five minutes assigning the next trials, each with a success criterion attached before anyone starts; five minutes on the running numbers. Short and boring beats long and inspirational here, because the meeting has to be cheap enough to survive a bad quarter.

    3. Mandate one agent for one workflow.

    An optional tools channel is where adoption goes to die, because optional means whoever is busiest that week opts out first and the rest follow. Pick one workflow, name it, and make the agent the default way that work gets done this week. People need one thing they are expected to use, not a menu.

    4. Set four things before any new agent goes past you.

    Decided upfront: the success metric in minutes or dollars, a trial window of two to four weeks, one named owner, and what you expect to save, written down before you start so you can check it against the result.

    5. Celebrate what works and cut what does not.

    Once a month, spend 30 minutes reviewing every agent in use. Keep the ones showing measurable savings that people are still running. Cut on three signals: nobody used it in four weeks, it creates more work than it saves, or it raises a security concern. Say the cut out loud in the standup and say why, so the team learns that retiring an agent is a normal outcome.

    Two product features carry part of the maintenance load once several agents are running. Evaluations are saved test cases that catch a regression before it ships, so a change to one agent's instructions cannot quietly break behavior that already worked, and Reflections have an agent review its own conversations on a schedule and propose improvements to its own instructions for you to approve.

    These five moves work in the order adoption changes, which is attitude, then habits, then actions. Attitude shifts when the team watches the thing work on your real task instead of a demo, and habits form on the standup and the monthly review.

    The ways a rollout stalls are predictable and most of them show up on the first build.

    See what usually goes wrong

    11 · When it breaks

    What usually goes wrong on a first build?

    The failure that shows up first on almost every build is an agent that produces something technically correct in a voice that is not yours, because it was never given a real example of your writing to work from. Scan the symptoms below for the one you are seeing; the cause and the fix sit underneath each.

    "It doesn't sound like me. It reads like a form letter."

    Cause: the agent was given a description of your voice instead of examples of it. Fix: build a voice Skill through the Upload Files path using three to five messages you sent, not messages written as samples. If a correction you make in chat produces the right voice, tell the agent to turn that into a Skill so the fix persists.

    Applies toInboxContent

    "It sent something instead of drafting it."

    Cause: the instructions described the intent without stating the prohibition, so nothing stopped the send. Fix: write the restriction as its own sentence in "Add instructions for the agent," in the form "Never send." Then back it up in the product rather than in prose. Gumloop's approval mechanisms include per-tool controls and App Rules, and an App Rule can block a specific call outright. A Skill can only ask; an App Rule stops the call.

    Applies toInboxContent

    "It made something up that isn't true."

    Cause: it was asked a question it had no source for and produced a plausible answer instead of an admission. Fix: attach real material as a Knowledge Source, since an agent with knowledge attached searches it and answers with citations. Then instruct it to write "not verified" when it cannot confirm something, and check the output for that phrase. A first run with no gaps anywhere usually means the gaps were filled rather than absent.

    "Nothing happens when it's supposed to run."

    Cause: the Trigger condition is wrong for the job, or a connected account's authorization expired. Fix: open the Triggers panel first and confirm you picked the right one of the three types, since a Scheduled Trigger will never fire on a new email and an Event-Based Trigger will never fire on a Monday morning. If the Trigger is right, re-authorize the Connector.

    Applies toAll five

    "I set it up and now I'm afraid to touch it."

    Cause: it went straight to a live schedule without a watched test period, so you have no evidence about how it behaves. Fix: pause the Trigger, run it manually three to five times on real inputs, and read every output completely.

    Applies toAll five

    "It sort of works, but I can't tell if it's right."

    Cause: the first build was scoped too large to verify in one sitting, so no single run tells you whether it worked. Fix: shrink the job until one manual run fits in an afternoon and you can read every output it produced. One task, one output, one owner. Scale comes from a second agent, not a bigger first one.

    Applies toAll five

    "It's pulling old or duplicate information."

    Cause: the lookback window is too wide, or nothing tells the agent what it has already handled. Fix: put an explicit date boundary in the instructions, in the form "only material published in the last seven days," and write a duplicate rule naming the field that identifies a record as the same one. For a data agent, that field is usually the email address.

    "My teammate ran it and got a different result than I did."

    Cause: the process lives in one person's head and only part of it reached the agent. Fix: this is a rollout gap rather than a build gap. Move the shared logic out of one agent's instructions and into a Skill, which is a reusable instruction pack anyone on the team can point an agent at, then give the agent one named owner. See team rollout.

    Applies toAll five, most often on teams
    Something else? Check the FAQ

    12 · Questions

    Frequently asked questions

    How do I start automating my business with AI?

    Pick one task you already do the same way every time and build one agent for that task. Give it something that starts it, connections to the tools involved, and a written job description. Watch it run on real work before you let it run on a schedule. Then use what you learned to build the second one.

    What should I automate first?

    Automate rule-based, repeatable, low-ambiguity work you already do the same way each time, because you can tell inside a single run whether the agent got it right. Skip anything that turns on a judgment call or on a relationship where the wrong tone costs you something. Common first builds: inbox follow-up, content repurposing, prospect research, admin data-entry, and a weekly research digest.

    What is a Gumloop agent actually made of?

    Six kinds of parts: a Trigger that starts it, Connectors that link it to outside tools, Skills that teach it specific procedures, Knowledge Sources it searches before answering, Subagents it can hand distinct jobs to, and Abilities like Web Search, Web Fetch, Image Generation, and a Code Sandbox. It is an assembly of parts you can open and read, not a model that already knows your business.

    When should I build a Skill instead of just writing instructions?

    Build a Skill when the procedure is long enough that you would otherwise paste a full document into the instructions box, or when a second agent will need the same procedure. Keep it in the instructions when the rule governs everything this one agent does and fits in a few lines. Gumloop's own test: if you are about to paste a long SOP into the instructions, make a Skill.

    When should an agent hand off to a Subagent?

    Hand work to a Subagent when it is a distinct job with its own tools, its own steps, and its own back-and-forth. The signal is Connector sprawl: if one agent is accumulating every integration you own, split it into specialists and call them as Subagents. An agent can also clone itself, shown as (Me), with a depth limit of one.

    Is Gumloop free?

    No. Gumloop runs on a 14-day trial, after which it moves to a paid plan. There is no permanent free tier. The one free thing here is the Agent Lab session at bnedai.com/agent-lab, where you build a working agent live with help. That session costs nothing to attend. The platform you build on does.

    What should an AI agent never be allowed to do without asking me first?

    Anything irreversible: sending something external, spending money, deleting data, or changing access and permissions. Put those behind a human yes. A Skill only asks an agent to behave; an App Rule stops the call at the tool.

    Why do AI automation efforts stall or die after the first few weeks?

    Because nobody owns the agent once the demo is over. The build works, the person who built it moves on, and no one is checking whether it still runs correctly. The fix is a named owner, a schedule, and a real check-in. The other common killer is a first build scoped too large to verify in one sitting.

    Next step: pick your first build

    13 · Next steps

    Where do I go from here?

    Pick one of the five recipes and build it this week, or bring the task to Agent Lab, the free live session where you build a working agent with help.

    Closest reads to this guide:

    Retake the assessment