Skip to main content
Almost everything that goes wrong at run time here is quiet on purpose: an unknown variable becomes an empty gap, a step the channel cannot do is skipped, a step with nobody to send to is skipped. Those are the right behaviours in front of a customer, and they mean a broken automation looks exactly like a working one. Two places say it out loud, and they are where most of this page ends: the editor’s checks panel before an automation goes live, and the run log afterwards. For agents it is the readiness checklist in the studio.

​
An automation did nothing

Work down these in order.
  • It is switched off. The card’s switch, and the pulsing dot beside the name, are the only thing that makes it live.
  • The trigger is not the one that happened. New message received is any incoming message; First message from a contact fires once per contact, ever; Keyword match needs the words to be present. Check the trigger pill on the card against what you actually did.
  • The channel scope excludes it. Selecting none means every channel; selecting some means only those. A rule scoped to Website never fires on WhatsApp.
  • A keyword trigger has no keywords. It can never match. Activation is refused for Keyword match and the checks panel flags it.
How to check. View runs in the card’s menu shows every run with its contact, channel and trigger. An empty list means it never fired, which is a different problem from a run that went wrong.Triggers · Diagnostics and logs
Twelve triggers carry no channel at all — the form, appointment, payment and store ones. A form is submitted on a public page and a payment is made on Razorpay’s, so there is no channel to compare against.What is happening. If you built the automation on a different trigger first and scoped it to a channel, that leftover scope used to sit invisibly on the rule. The picker is now hidden for these triggers and any leftover scope is cleared, and the engine ignores channel scoping for them entirely.What to do. Re-open the automation, confirm the trigger, and save. If you built it on a store trigger from the store console it is tied to that store; one you build by hand fires for every connected store.Triggers · Channels and the window
A scheduled automation runs for the workspace, not for one person. There is no contact and no conversation, so every step that messages, tags or notes somebody has nobody to act on, and the editor marks those steps as unable to run.What to do. Use a schedule for steps that need no contact — an HTTP request, a Google action, a team notification. To reach people on a timer, use a broadcast instead.{{ message.text }} in a scheduled automation is flagged by the checks panel for the same reason: there is no message.Triggers · Variables
Turning Active on is validated by the server, and a broken automation is refused with the first problem it finds and where it is.What it requires. At least one step, and each step carrying what it needs: message text, a template name, a tag, a pipeline and stage and title, a valid URL, a wait greater than zero with a real unit, a split percentage between 0 and 100, at most 3 buttons, at most 10 list rows, a variable name of letters, numbers and underscores. Keyword match needs a keyword, On a schedule needs a schedule, Tag added needs a tag.What to do. Open the checks panel before you try — it lists everything, and clicking a finding selects the step it is about.Diagnostics and logs

​
It ran, but not the way you expected

Skipped is not a failure. It means the channel could not do that step, there was nobody to send to, or the step is switched off.The usual cases. A WhatsApp template or a list on Instagram. A sending step on a run started by a hosted form, a booking, a payment or a store order, where the person may never have messaged you and there is no thread — the tag, segment, deal and notify steps beside it still run, and the run ends partial.What to do. Read the run log: it names which step was skipped and why. To behave differently per channel, add a Condition on the Channel subject and put the template send in one branch and plain text in the other.
Send an invoice is the exception among the sending steps: it creates the document anyway and records that it could not deliver it.
Channels and the window · Diagnostics and logs
An unknown token resolves to an empty string. It is never shown to the customer with its braces intact — a visible token reads as a broken app — but it does mean a typo produces a gap rather than an error.The four causes. A misspelled path. A token pointing at a step that runs after this one, or on the other side of a branch — it could only ever be empty. A namespace that is not a real one. A variable no Set variable step and no Also save the result as ever writes.What to do. Use the insert data button beside the field rather than typing a path: it lists only what this step can actually read, with a value from a recent run beside it. The checks panel flags all four. {{ contact.name | default:"there" }} avoids “Hi ,” when a name is genuinely missing.Variables and expressions · Diagnostics and logs
A {{ that is never closed is sent to the contact exactly as typed — there is nothing for the engine to resolve.What to do. Close it. The checks panel reports an unclosed brace before you go live.Variables and expressions
A parked run does not stop WhatsApp’s clock. After a wait, only an approved template will send unless the customer has messaged again in the meantime.How to check. The checks panel adds up every Wait along each path — counting a Wait until as up to 24 hours — and warns on the first sending step that lands past the limit, naming the total.Two fixes. Shorten the wait, or make that step a Send template. The same applies to Request a payment: outside the window the link is carried by an approved template and the step’s own message text is not used, because a template’s copy is fixed at approval.Channels and the window · The 24-hour window
Engines do not know about each other unless there is a rule saying so, and there are only two such rules.
  • An automation beats an AI agent. If an active automation on a message trigger, scoped to this channel, with matching keywords, answers the message, the agent stays quiet on it.
  • A comment funnel beats the Instagram comment trigger on the comment it claims.
Everything else can double up. Two automations listening for the same keyword both fire — the trigger panel says so when you add a word another automation or a flow already listens for. A flow and an automation can both answer the same WhatsApp message.What to do. Give each keyword one owner, or narrow one of them with a channel scope or a stricter Is exactly the keyword match.Triggers · Why is my agent silent?
A step set to Carry on to the next step on failure still marks the whole run failed, so a broken step is never invisible. The rest of the run still happens.What to do. Expand the run and read the step that failed; its reason is on the row.Diagnostics and logs
Only an active flow answers anything. A draft flow is skipped, including by an automation’s Start flow step, so the step runs and nothing happens.What to do. Publish the flow. The checks panel flags a Start flow pointing at a flow that is not active, and a Run automation pointing at itself or at an automation that is switched off.Flows · Diagnostics and logs
Requests to private, loopback and internal addresses are refused by the server. The checks panel warns about them, and about plain http://, before you go live.What to do. Point it at a public https:// address. The step publishes {{ steps.<name>.status }} and {{ steps.<name>.body }} even on a refused call, so a Condition after it can branch on the failure instead of the automation simply stopping.Steps · Variables and expressions
Only static segments have a membership anything can be added to or removed from. A filter segment works its membership out on the spot from its rules, so there is nothing to write.What to do. Point the step at a static segment. The checks panel flags a segment step pointing at a filter segment; whether the segment still exists and still belongs to this workspace is checked when the step runs, because that answer can change after activation.Segments · Steps

​
Flows

Flows are WhatsApp only. They send lists, approved templates and catalogue products, none of which exist on the other two channels, so an Instagram DM or a web chat never enters the flow engine at all.What to do. Build it as an automation instead — that is the channel-agnostic engine, and it skips the individual steps a channel cannot do rather than failing mid-conversation.Flows
A contact can be in one flow at a time. That is why Connect flow ends the current run before starting the target rather than nesting, and why a loop between two flows terminates instead of growing.Check. Open Runs in the editor toolbar. An Active run shows which node it is sitting on. A run that has gone quiet is closed by a sweep — the default cutoff is 24 hours without progress — so it does not block the contact for ever.Also. If somebody on your team replied to that contact by hand, the run is marked paused by agent and stops advancing. A person owns the conversation now.Waits, hand-offs and runs
The customer’s reply did not match what that node was waiting for — they typed a sentence instead of tapping a button, or sent a photo when a document was asked for. The flow re-sends the same prompt, by default twice, and then hands the conversation to a person.What to do. Expand the run and read the event timeline: fallback_fired is the re-prompt and handoff is the hand-over. Widen what the node accepts, or make the prompt clearer about tapping rather than typing.
The text of a customer’s reply is not written to the event log — only its length, and which button or row was tapped.
Waits, hand-offs and runs · Node types
The link could not be created at all — Razorpay is not connected, or the keys were revoked. There is nothing to wait for, so the run takes the If they don’t, go to path immediately rather than parking for ever.What to do. Connect or reconnect Razorpay and test the keys. When a link is created normally, the run parks until the link’s expiry, and a payment that arrives first claims that parking and continues down When they pay, go to — exactly one of the two wins.Razorpay payments · Waits, hand-offs and runs
The validation bar at the foot of the editor checks the graph from the entry node. A node nothing points at is listed as a warning — it does not stop activation, but nothing will ever run it.The related error. A path pointing at a node that is not there, or a required path left empty, is an error and stops activation. Clicking either jumps to the node on the canvas.What to do. Connect the node, or delete it. Then check the run history for where real runs actually stop — a run that reaches a path pointing nowhere is marked Failed.The flow editor · Waits, hand-offs and runs

​
AI agents

An agent that is silent looks the same from the inbox whatever the cause, so there is a list of causes in the order they are checked.Start with the readiness panel above the studio tabs. It names anything outstanding and each line takes you to the tab that fixes it. An amber panel means a blocker: the workspace cannot reach a model, or the agent is switched off.Then the run-time gates, in order: something else already handled the message (a flow, a button tap, a message with no text); no switched-on agent covers this channel; the agent does not answer automatically; the workspace is out of credits; a teammate owns the thread; AI replies are paused on that conversation; the conversation’s reply budget is used up; test mode excludes this customer; an automation already answers this message; a handoff phrase matched; the model handed off or came back empty; the provider call failed.Two things are deliberately not blockers. Drafts only and test mode are legitimate ways to run an agent.Why is my agent silent? · Routing
Three limits end a thread rather than a message.
  • The reply budget. Max replies per conversation — three by default — counted over a rolling 24-hour window that opens with the first reply after a quiet period. A customer who comes back the next day gets a full budget.
  • A handoff. The agent decided it could not help, or a handoff phrase matched. AI replies are paused on that conversation and it waits for your team.
  • An owner. A conversation assigned to a member is that member’s, and the agent stays out of it.
What to do. The bot button in the thread header resumes AI replies on that one conversation, and clearing the owner lets the agent answer again.Why is my agent silent? · Working a conversation
Routing is decided once per conversation, not per message. The first AI reply latches that agent to the thread and every later message goes to the same one — otherwise reordering the list would change who the customer is talking to mid-sentence.For a new conversation, the first switched-on agent in list order whose channels cover it answers. Selecting no channels means every channel. There is no fallback: an agent scoped to Instagram will not answer a WhatsApp thread just because it is the only one you have.What to do. Reorder the list with the arrows on each row — the list is the routing order — and put the narrow, specific agent above the general one. To move one existing thread, use the owner control in the thread header.Routing
The reply reads as though the agent never knew it could book a meeting, log a row or send a basket link. That is a withheld tool: it was never shown to the agent, so it could not mention it.Every writing tool is withheld while drafting, deliberately: pressing the draft button three times must not book three meetings or send three payment requests.Skills · Why is my agent silent?
Open the test drawer and ask the same question. The evidence line under the reply names the documents the answer was built from and every tool that ran.
  • The right document was not retrieved. Split a long document into one topic per document, and check its status is Ready rather than Keyword only, Needs reindex or Failed.
  • It was retrieved and ignored. A ground rule or an instruction is contradicting it — ground rules outrank everything below them, including your documents.
  • It invented something. Fill in What the business does and add the fact as a document. The agent is told not to guess, but with nothing to go on it has only the customer’s question.
  • It sounds wrong rather than being wrong. That is tone, reply length and Extra voice notes on the Persona tab.
Knowledge · Persona, voice and behaviour
Every document carries a status, because a document that saved but is invisible to the agent is the worst outcome here.Also check the selection. An agent set to Only these reads only the ticked documents, and documents added later are not included automatically. Selecting nothing means exactly that.Knowledge
At zero, the inbox draft button refuses with a message offering a top-up, and the auto-reply agent leaves conversations for a person rather than telling a customer a business is out of credits. Nothing is lost; the conversation waits.What to do. Top up from the header badge or Provider & credits, or save your own provider key. Fallback only ever runs one way: with credits exhausted and your own key saved, your key is used — but a working own-key setup never falls back to spending credits.What spends them. Auto-replies, the inbox draft button, the test drawer, indexing documents, and the AI builders for automations and flows.Credits and your own key
Test mode is a phone-number allowlist. Only the numbers you listed are answered — and Instagram and website conversations have no phone number, so they are never answered while it is on.Check. The numbers are matched exactly, with the country code: +91 98765 43210. A Test mode pill sits on the agent’s row and in the studio header so it cannot be left on by accident.What to do. Switch it off to go live. Clearing the last number also turns it off.Test mode and the playground
Nothing is sent to a customer from the drawer, and tools that read one customer’s data say so instead of reading somebody’s real orders. But tools that write still write: a Google booking, a spreadsheet row or one of your custom actions called from the drawer does the real thing.The drawer is also metered exactly like production — an unmetered test surface would be an open AI endpoint behind a login page.Test mode and the playground
How many agents a workspace may have is a plan limit: Starter 1, Growth 5, Enterprise unlimited. A paused agent still occupies a slot — a cap you can dodge by pausing is not a cap.What to do. Delete an agent you no longer use, or change plan. The meter under the page title says where you are.Create an agent · Plans and billing

​
Still stuck