How to create and activate an AI agent

Before you start

Agents is visible to every signed-in account, with no invitation needed. On Free, open Agents and choose Create agent to see upgrade options. Creating and running agents requires Starter or Growth and available AI credits. Create one focused agent first. Add another only when the next job needs different rules, knowledge, or confidence settings.

Check your agents’ activity

1. Open AI → Agents to see active conversations now, confirmed agent messages sent in the last 7 days, and items awaiting review. Messages sent counts each chat bubble; Playground and failed sends are excluded.
2. Choose an agent’s name to edit it, or Test to open its Playground. Last active shows its latest conversation-session activity. An awaiting-review count opens that agent’s first loaded pending item; a + means more reviews may be available.
3. Use the compact Review queue below the list to read and approve pending replies. Open queue shows the full queue.
4. Choose one of the five Recent conversations to open it in Inbox. Hidden and deleted contacts are excluded. If Knowledge needs attention appears, choose Review sources to inspect and retry failed indexing.

Generate the agent prompt

1. Open AI → Agents and choose Create agent.
2. Choose Generate with AI at the top. Use the suggested jobs for ideas, describe the job, voice, and rules, then choose Generate instructions. The optional Drafting options choose the model and billing source, which carry into the new agent.
3. Review the instructions, name the agent, and choose Continue to knowledge. Set up myself is also available at the top of the initial form if you want to write or paste your own instructions.
4. In Knowledge, choose Upload files or Write text to add knowledge, or Add existing to select saved sources. New sources are saved to your library immediately. Choose Continue to resources when ready; you can also continue without selecting sources.
5. In Resources, choose the links, products, workflows, contact fields, tags, and actions the agent can use. You can leave these empty and add them later. Back preserves your instructions and selections. Choose Create agent to save everything and open the editor. New agents are inactive and start with Review first.
6. In Instructions, describe when to share links, collect an email, recommend a product, or run a workflow, including when to stop and what to do next. In Behavior → Direct messages, choose All inbox messages, Keywords only, Intent or topic, or Manual only. Intent/topic routing requires a confident match; a configured topic must also match. Existing conversation controllers take precedence. Knowledge holds the saved sources and Resources and actions permissions; wait for sources to finish indexing before testing. Existing active goal and procedure rules appear in Instructions; there are no separate editors for them.
7. For public replies, open Behavior → Public comments and enable Reply to public comments. Choose Reply, Send to review, or Ignore for each intent. Start with Review every comment reply. Public replies cannot collect email or start a DM.
8. Settings → Model and billing holds the saved provider and model. Settings also holds reasoning, confidence, and follow-up limits.
9. Use Test agent beside the editor or open it on mobile. Select Direct message or Public comment, try realistic messages, and inspect the detected intent and reply details. Changing the channel starts a new saved test conversation. No customer messages or real actions run.
10. Save draft keeps the current live version running. Choose Review & activate to publish and start an inactive agent, or Publish changes for an active agent. The dialog summarizes the reply mode, channels, and sources. Pause takes effect immediately.

Start manually instead

1. Open AI → Agents, choose Create agent, then choose Set up myself at the top.
2. Enter a practical name such as Coaching helper or Program questions, write or paste your instructions, then choose Continue to knowledge.
3. Add knowledge from saved sources, uploads, or pasted text, then choose Continue to resources. Select the links, products, workflows, and actions the agent can use, or leave either step empty to finish it later. Choose Create agent on the final step. The editor opens with your instructions and selections saved; Review first is already selected under Direct messages.
4. Write or paste your instructions. Generate instructions proposes a fuller document from your input, using Everyday, Advanced, or Expert with default reasoning. Paste an existing prompt directly into Instructions without spending AI credits; business facts still belong in Knowledge. If you close the generation dialog with input or a generated proposal, confirm to discard it or choose Cancel to keep working. Applying a proposal never changes permissions or activation. The editor and its Generate with AI toolbar helper support up to 500,000 characters for agent instructions. An incomplete or over-limit AI result shows an error and leaves the document unchanged.

Set reply timing

1. In Behavior → Direct messages, open the reply-mode settings and find Reply timing.
2. Assigned agents wait for 10 seconds without a new incoming message before generating an automatic reply or Review draft. Each new message restarts the wait and replaces older pending drafts. If a message arrives during generation, that answer is discarded and the next generation waits for the sequence to settle. Explicit AI Reply requests keep their existing behavior.
3. Choose Reply right away, Same wait every time, Random wait, or Set by reply number. Random wait picks a fresh time between At least and At most for each automatic reply.
4. For a sequence, set First reply to 2 minutes and Second reply to 5, then Every reply after that to 3. Use Add reply wait for more steps, up to 10. Each wait accepts whole minutes from 0 to 60; 0 means right away.
5. Save and publish. The next reply run uses the new timing; a reply already in progress keeps its snapshot. The sequence counts successfully sent replies and resets in each new session.
6. Timing applies to automatic Inbox and workflow agent replies. Playground and manually approved replies skip the wait. Follow-ups have a separate schedule. New incoming messages replace waiting replies, and taking over cancels them. Delivery can take a little longer than the chosen wait; session expiry and provider messaging windows still apply.

Break long replies into chat messages

1. In Behavior → Direct messages, turn on Break up long replies below Reply timing. It is off by default.
2. Choose a maximum of 2 to 4 messages and a pause of 1 to 5 seconds. The defaults are up to 3 messages with 2 seconds between them. Short replies can stay together.
3. A second AI step splits the completed answer without rewriting it. With LinkStack credits selected, each response with text costs 1 extra AI credit, including Playground, even if the result stays in one message or formatting fails. Silent turns add no breakup credit. If formatting fails, the original reply stays together.
4. Save and publish to use the setting on the next reply run. Test agent previews all messages immediately and includes the extra step in the credit total. In-progress replies keep their settings.
5. Reply timing delays the first message; the short pause separates the rest. The group counts as one agent reply. New inbound messages, taking over, expired sessions, or provider restrictions stop further messages. Smart Link buttons appear on the final message.
6. Review Queue previews the saved messages before approval. If you edit, use blank lines to separate messages up to the chosen limit; extra paragraphs stay in the final message. Editing and sending saved text add no generation charge. When approved actions need a final AI reply based on their results, that pass is metered once and changed text needs another approval. A partial send is not automatically replayed: inspect Inbox before sending any missing text manually.

Teach voice and check behavior

1. In Instructions, describe the voice, reply length, emoji use, and phrases to use or avoid in your own words. Put examples under an Example replies heading. These details stay in the same document; there are no separate voice sliders or example fields.
2. Paste a few of your own replies under Example replies, and explain which parts of the wording the agent should follow.
3. Use Test agent to try nice comments, questions, complaints, and link requests. Inspect reply details and Improve AI feedback, then adjust Instructions, Behavior, or the sources and resource access in Knowledge. Save and publish when ready.

Send and track link buttons

1. In Knowledge → Resources and actions, choose Links and select the saved Links the agent may share. Turn on Send saved links below the list, save, and publish. Use Instructions to describe when it should offer each link.
2. The agent attaches a labeled button to its message. Live sends automatically add UTMs for the source, DM channel, agent campaign, and a unique send code. You do not need a separate saved Link for each recipient, and names, emails, and usernames stay out of the URL. Product buttons receive the same UTMs on their public URLs. Playground sends no customer messages and creates no live tracking code.
3. Open the saved Link's analytics and scroll to Agent campaigns. Creators are grouped by agent, with names and @handles when available, channel, sends, clicks, last sent, and last clicked. Select View contact to open their Audience profile. Recipients with no clicks still appear. Counts follow the selected date range, and clicks can come from earlier sends. Use the page controls to see more recipients.
4. Click feed lists individual clicks with a Contact link when a match is available. Not matched means the click has no available contact match. Older untracked sends and clicks are not matched retroactively. Hidden or deleted contacts are not named. In Inbox, the recipient's contact sidebar also shows Latest link click. Crawler previews and test clicks are excluded. Product-page visits keep their existing product reporting.
5. If Instagram rejects the button format, LinkStack sends a short message link with the same tracking. Forwarded links still attribute clicks to the original recipient's send; a click is not proof of who opened it.

Read current products and verified customer records

1. In Knowledge → Resources and actions, choose Products to select the published LinkStack products this agent may read and recommend. Turn on Recommend products below the list. Save and publish. Product reads use current facts and server-resolved public destinations.
2. Enable Read verified purchases and access only when the agent should handle customer-specific purchase or access questions for those products.
3. Open the customer's Audience profile. Save their email, then use Customer identity for AI to confirm their identity and record how you independently established it. A claimed email or matching order alone is not verification.
4. Revoke confirmation to remove access immediately. Changing the saved email or reassigning the conversation also removes the old confirmation. Test verified and unverified customer fixtures in Playground before activating the agent.

Configure handoffs and follow-ups

1. In Knowledge → Resources and actions, choose Links, Products, or another category to search and select saved resources. Use the permission switches below the list, or open Permissions for all actions. Selecting a resource does not enable its action. Counts update immediately, and choices stay in your agent draft across categories and tabs. Check a folder to select its current contents and subfolders; new items added later are not automatically approved. Use Selected to review, then save and publish. For a selected workflow, open Configure to set its label, required fields, and Call and return or Permanent handoff. Apply selection keeps configuration edits; Cancel discards them. Put when to use the workflow in Instructions.
2. Describe task steps, destinations, next actions, and when to stop asking in Instructions. Intent routing and public-comment rules are in Behavior. Per-task limits in Instructions guide the model; approved actions and resources remain enforced by the app.
3. In Settings, open Follow-ups and session limits to opt into scheduled follow-ups. Set the delay, maximum, timezone, and quiet hours. Write their purpose and wording guidance in Instructions.
4. In Activity → Recent sessions, identify the person and message, then expand the row. View conversation opens its Inbox thread in a new tab; Reply details opens the trace in that row. Show more sessions reveals older rows. Use Take over to stop AI control. A manual Inbox reply asks whether to keep the agent or turn it off; Remember my choice saves the selection. Reassign a published active agent explicitly to resume.

Resolve unanswered questions

1. Open Activity → Knowledge gaps. Only live replies marked unable to answer enter this list; low confidence alone and Playground tests do not. Matching questions after case and punctuation normalization increase the count.
2. Write an accurate answer and choose Save answer and resolve. This creates a knowledge source and clears the row. Check its indexing and source selection before testing and publishing the agent. Resolved questions do not automatically reopen.

Safe defaults

• New agents start in Review first, including AI-generated drafts. Keep it selected to approve replies and actions before they run.
• Use Keywords only for specialized agents.
• Keep Manual only when the agent should stay out of automatic inbox routing.
• New agents have Time limit set to Off. To limit Inbox conversations, open Settings → Follow-ups and session limits and choose a duration, or enter a custom duration in minutes, hours, or days. Limits start at assignment and do not reset on replies. Workflow steps have their own Time limit setting. Saved explicit limits and sessions already running keep their existing deadlines.
• Move to Auto-send when confident only after tests and reviews are consistently right.

Find an archived agent

1. Open AI → Agents and choose Show archived beside Create agent. The control is also available when the usual list is empty.
2. Find the agent with an Archived status and choose Open to inspect its saved setup. Deleting an agent archives it and keeps its setup history.
3. Choose Hide archived to return to the usual list. Showing archived agents does not activate them or count them toward your plan's agent limit.