Skip to main content

Agent objectives

Module goal: set up machine-checkable objectives so that the agent does not re-ask for what it already has.
The Agent objectives window with two objectives

Agent objectives (Quality & integrations → Agent objectives). Two objectives of the tutorial agent: “Visit confirmed” (required) and “SMS sent” — optional, available only after the first one.

The Agent objectives window with two objectives
The problem this solves is the following: the agent does not remember the whole conversation, only the latest replies. On a long call, an early reply where the person already gave their name falls out of that memory — and the agent asks a second time. Asking “do not re-ask” in the prompt does not help here: the agent does not forget the rule, it forgets the fact. An objective is a fact of the conversation expressed as a condition over variables: {{confirm_text}} != "". The code marks it, not the model: the model only reads the objective state and never writes to it. This gives properties a prompt cannot:
  • the agent does not re-ask about a closed objective — the summary of what is established goes to the tail of the prompt and is refreshed on every turn, so it never falls out of memory;
  • an objective can be made revisable — the customer changed their mind, the cancellation condition fired, the objective is open again;
  • preconditions set dependencies: sending an SMS reminder only makes sense after the visit is confirmed;
  • transition formulas can use the goals_completed() predicate — “everything important is established, time to say goodbye”.
Objectives are set up in the agent settings: Quality & integrations → Agent objectives. The first action in the section is the Track agent objectives switch in its header. While it is off or the list is empty, the mechanism is fully disabled and has no effect on the agent’s behaviour. Edits go to the draft with the window’s Save button.
Agent objectives are not Goal achieved. The Goal achieved figure in campaign statistics (Goal reached on the dashboard) is a post-call analytics evaluation: after the call, the model decides whether the call succeeded by the agent’s success criterion (module 18). Agent objectives work during the call, are marked by code from a condition, and do not affect that figure.
Two things people often ask. First: the summary does not appear in the conversation transcript — in the call record you see only the real replies. Second: the mechanism makes no extra language-model calls, so it does not affect call latency. And one more: there is no need to copy the summary into node prompts by hand — it is mixed in automatically, and a hand-written copy quickly drifts away from the settings.

How conditions are set

The Completion condition and the Cancellation condition are assembled with the same condition builder that works on transitions: you pick a variable, an operation and a value, and the cards add up to an expression. A condition can also be edited as text — the Edit as text button. The Node visited check is added as a separate card (the + Add visited node button): the node is picked from a list. It is the only value in an objective condition that the author has nowhere to look up — so it is picked from a list rather than typed in. A moment after an edit, the line under the condition shows Checking…, then Condition parsed or Condition is not valid with the reason. The wording of an objective — the Established fact field — is a statement of fact (“identity confirmed”), not a task (“confirm identity”): the text goes to the agent, and a task-shaped text ends up being read out loud — “right, I still need to find out why the payment was missed”. Proofread the wording on test calls with the same care you give the agent’s replies.
Tip: the language of the wording. Write the established fact in the language the agent works in. For an English-speaking agent under the US profile (us_tcpa): “The customer confirmed the appointment or picked a new time”; for a Russian-speaking agent: «Клиент подтвердил визит или выбрал новое время».

Condition cheat sheet

Comparisons are case-insensitive: "Yes" and "yes" are the same. Expressions with brackets or a mix of AND and OR are edited as text only — the cards do not show that form.
Tip: values in the conversation language. If a field is filled from the customer’s answer, compare it with the values that actually arrive in calls. For a bilingual agent, check both variants: {{consent}} == "yes" OR {{consent}} == "да"; for an English-speaking agent (us_tcpa): {{consent}} == "yes", {{intent}} CONTAINS pay.

Where field values come from

This is the main practical question of the setup. There are three sources:
  1. Extraction from the conversation — what the agent finds out from the person: a name, a date, a reason, an answer to a question. In free mode extraction runs on every turn, in a rigid flow only on the steps where it is configured.
  2. Results of calls to your systems — every request automatically exposes {{name}}_success, {{name}}_status, {{name}}_error plus the fields you mapped from the response. A useful side effect: if the same request is repeated and returns a failure, the objective stops being closed by itself — no separate cancellation condition is needed for that.
  3. Data known before the call — call and campaign variables: the name from the contact file, the debt amount, the contract number.
A delay of exactly one reply. Objective state is recomputed once at the end of each turn, so an objective closed by the answer on the current reply enters the summary and the goals_completed() flag starting from the next turn. This is normal, not a fault.
Pitfall: an objective on a variable nobody fillsIf no node fills the condition’s variable, the objective will never close — validation warns about it but does not forbid it: the variable may legitimately come from the campaign or the API. Check that it really arrives.
The draft keeps an objective with an unparsable condition — but it blocks publishing the agent and making a test call. So an unfinished condition does not stop you from putting the work off until tomorrow, but it will not slip into production unnoticed either.

What is visible in the call history

The History section → call card → the Transcript tab with the Show system events checkbox on (administrators and developers have it). The timeline shows three events: Objective closed (showing on which reply), Objective reopened (with the reason — the cancellation condition fired, the completion condition no longer holds, or a precondition was retracted) and Objective condition error. If a condition breaks during a call — a variable was renamed, say — the call is not interrupted: the objective counts as not closed, the warning is written once, and the conversation goes on normally. More on the card: Call details.

The “everything established” flag ≠ call success

The goals_completed() flag is raised once every objective marked required is closed — optional ones do not hold it back. The call does not end by itself: a conversation rarely ends on the last checkbox, the person still says goodbye and asks about the date again. Where to route the flag is up to you — to a farewell block, to a handover to an operator, to a final confirmation. The call success evaluation from post-call analytics — the one that goes into the campaign’s Goal achieved — is computed independently and means something else: “everything planned was established” versus “the call succeeded by your criterion”. Both are visible in the record.

What objectives do not do

  • They do not work with speech-to-speech (realtime) agents — such agents do not fill variables as the conversation goes, and without them there is nothing to check. The switch can be turned on, but it will have no effect.
  • They do not replace mandatory disclosures. A call-recording notice, identification before discussing a debt and, under the us_tcpa profile, telling the person that an AI agent is calling are not objectives: an objective hints to the agent, and a hint can be ignored. Where something must be said, the flow needs a hard route.
  • They do not impose order — only preconditions, and an unavailable objective is not shown to the agent at all, so it does not raise a blocked topic too early. An answer given early is not lost: the value is kept, and the objective closes the moment its precondition is met.
  • In a rigid flow an objective on a variable the flow never fills will not close — free mode has no such limitation.
  • Objectives live on the agent and are versioned with it.
The full reference for objective fields and blocking checks is on the Agent objectives page.