# Best practices for reliable voice agents > Build reliable HeyRik voice agents with scoped jobs, grounded knowledge, explicit variables, safe handoffs, realistic call testing, controlled launch, and continuous improvement. Source: https://www.heyrik.com/docs/academy/best-practices Use specific instructions, grounded facts, observable outcomes, controlled handoffs, realistic testing, and a disciplined improvement loop to keep an agent useful after launch. 18 minute guide Production readiness checklist Reliability is not created by a clever prompt alone. It comes from a narrow job, explicit behavior, approved facts, constrained tools, observable outputs, realistic testing, and an operator who reviews what actually happened after launch. **Make every rule testable** “Be helpful,” “sound human,” and “do not hallucinate” cannot be scored consistently. A good rule names the trigger and the expected action—for example, “If approved knowledge does not contain the price, say it requires confirmation and offer a human follow-up.” ## Give one agent one primary job A focused scope reduces prompt conflicts, shortens calls, simplifies testing, and creates an outcome that can be measured. ### One audience Name who the agent is speaking with and what they are expected to know before the call begins. ### One completion condition Define the observable moment the job is done: confirmed requirements, an approved answer, or a tool-confirmed action. ### One boundary Name requests this agent must not handle and the transfer, follow-up, or stop behavior for each. ### One useful result Capture only the output fields the next person or system will use to make a decision. Avoid ### Unmeasurable scope Be a helpful sales and support assistant that answers everything, promotes products, books meetings, handles complaints, and keeps customers happy. Recommended ### Measurable scope For inbound property enquiries, capture and confirm property type, budget, location, timeline, and permission for a human follow-up; never state unverified price or availability. ## Design for a real spoken conversation Phone callers pause, interrupt, correct themselves, refuse questions, change language, and ask unrelated things. The flow must support those behaviors without losing its job. ### Turn length Instead of Ask four questions in one long paragraph. Use Ask one short question, listen, acknowledge, then ask the next relevant question. ### Confirmation Instead of Silently store names, amounts, dates, and times. Use Repeat important values in a natural sentence and ask whether they are correct. ### Correction Instead of Continue using the first answer. Use Acknowledge the correction, replace the value, confirm it, and resume from the right step. ### Clarification Instead of Repeat the identical question louder. Use Rephrase once with simpler words or a small number of concrete options. ### Refusal Instead of Keep asking until the caller answers. Use Skip a nonessential field; for a required field, explain why it is needed and offer the safe alternative. ### Closing Instead of Say “thanks, bye” as soon as fields are filled. Use Summarize confirmed details and the real next action, invite correction, then end. **Do not confuse a script with a conversation** Write goals, order, branches, and constraints—not a rigid line-by-line performance that breaks as soon as the caller answers unexpectedly. ## Separate behavior, facts, and actions Put each kind of information in the layer designed to maintain it. Mixing them creates stale facts, hidden permissions, and prompts nobody can review. **Instructions** Identity, tone, sequence, decisions, limits, recovery, transfer, and closing behavior. How the agent behaves **Knowledge Base** Approved prices, policies, hours, locations, products, services, and frequently asked questions. What the agent may state **Integrations** Current lookups and real actions such as checking an order or creating a calendar event. What the agent may fetch or do **Call Settings** Speaking, turn-taking, silence, interruption, transfer destination, hard limits, and post-call delivery. How the channel behaves **0 of 6 complete**Loading saved progress… **01**Every factual source has an owner and review date outside the prompt.**02**The agent is attached only to sources relevant to its job.**03**Conflicting, obsolete, internal-only, and unsupported content is removed before upload.**04**The prompt says what to do when knowledge has no answer.**05**Each tool has a narrow description, required inputs, authorized action, success signal, and failure fallback.**06**The agent never describes an action as complete until the configured tool confirms success. ## Treat variables as a data contract Input and output variables connect a natural conversation to the rest of the workflow. Vague names and descriptions create silent operational errors. **Weak variable definitions** ``` Input: name Default: — Output: budget Description: The budget Output: qualified Description: Whether qualified ``` **Specific variable definitions** ``` Input: customer_name Default: there Description: Name supplied before the call Output: budget Description: Confirmed amount or range with currency; otherwise unclear Output: qualification_status Description: qualified, unqualified, follow_up_needed, or unclear ``` ### Safe input defaults A missing value must produce a natural sentence or disappear cleanly; a literal placeholder must never be spoken. ### Finite output shapes Use explicit outcomes such as yes, no, or unclear when a downstream system depends on consistent values. ### Minimum necessary data Do not capture sensitive or unrelated information merely because it appears in the conversation. ### Confirm before capture Important values should be repeated and confirmed so the transcript and structured result agree. ## Write the limitation and the next action together A refusal without a useful next step frustrates callers; a next step without a limitation encourages the agent to overreach. 01 ### Missing fact Say the information is not available in approved sources, record the question, and offer human follow-up. **Avoid:** Answer as accurately as possible. 02 ### Human request Transfer when enabled and the condition matches; otherwise explain that transfer is unavailable and offer the approved follow-up. **Avoid:** Try to retain the caller before transferring. 03 ### Sensitive request Decline to collect passwords, one-time codes, full card data, or unrelated personal information and direct the caller to the approved secure path. **Avoid:** Ask for whatever is needed to help. 04 ### Regulated advice State the permitted factual scope and route medical, legal, financial, or other regulated judgment to an authorized person. **Avoid:** Add a disclaimer and continue giving advice. 05 ### Action failure Say the action was not completed, preserve confirmed inputs, and offer the configured retry or human path. **Avoid:** Tell the caller it will probably happen later. 06 ### Opt-out or wrong person Apologize, capture the correct outcome, and end without continuing the flow. **Avoid:** Ask a few questions before ending. **Test the actual handoff** A written transfer rule is not enough. Verify the destination, caller experience, failure behavior, hours, and what context the receiving person receives. ## Tune for comprehension before personality A pleasant voice cannot rescue clipped turns, long speeches, poor silence behavior, or a language mismatch. ### Voice choice Instead of Select by name, avatar, or novelty. Use Test comprehension in the target language on a phone speaker with business-specific terms. ### Speaking speed Instead of Make every sales agent fast and energetic. Use Begin near natural speed and adjust gradually after listener testing. ### Turn ending Instead of Use the shortest possible pause. Use Allow normal thinking pauses without making every exchange feel slow. ### Interruption Instead of Disable interruption to protect the script. Use Let callers interrupt where appropriate and verify that the agent resumes with context. ### Silence Instead of Wait indefinitely or hang up immediately. Use Check in after the configured delay, retry a limited number of times, then close politely. ### Maximum duration Instead of Rely on the prompt to end eventually. Use Set a hard cap appropriate for the job and still write a natural closing. **Voice clones require continuing governance** Keep proof of speaker consent, limit who can assign the clone, review the current credit price before creation, and remove access when permission or business need ends. ## Use a repeatable test suite A test should state the setup, caller behavior, expected agent behavior, expected captured data, and pass or fail result. **Happy path** Caller gives valid answers in the intended order. Baseline completion **Out of order** Caller provides later information early or changes topic. Retain, do not re-ask **Correction** Caller changes a confirmed value. Replace and reconfirm **Ambiguity** Caller gives an unclear amount, date, location, or intent. Clarify once **Missing knowledge** Caller asks for an unsupported fact. Limit and escalate **Tool failure** Integration errors or returns incomplete data. No false success **Silence/interruption** Caller pauses or speaks over the agent. Natural recovery **Refusal/opt-out** Caller declines a question or future contact. Respect and stop **Human request** Caller explicitly asks for a person. Correct handoff **Unsafe request** Caller requests prohibited or sensitive action. Boundary plus next step **Score the result, not the vibe** Record whether the expected behavior occurred, whether the transcript supports the structured outputs, and where the first divergence happened. “It sounded good” is not a passing result. ## Operate the agent after launch Production traffic changes the evidence, not the responsibility. Monitor early calls, keep rollback available, and separate factual fixes from behavioral fixes. 1 ### Start small Use controlled inbound traffic or a very small outbound set before sending broad customer traffic. 2 ### Review the first calls Read transcripts, outcomes, captured fields, recordings when available, and tool invocations—not only aggregate metrics. 3 ### Classify the failure Choose prompt, knowledge, variable, integration, voice, call setting, channel, or operator process as the likely layer. 4 ### Change one cause Make the smallest change that addresses the earliest divergence and preserves behavior that already passed. 5 ### Run regression tests Repeat the failed scenario and the entire critical test set before restoring or increasing traffic. 6 ### Keep rollback ready Return the agent to Draft, unassign the channel, pause a campaign, or route to a person when risk exceeds the launch threshold. ### Wrong fact Instead of Make the prompt more forceful. Use Correct the knowledge source and test both present and absent answers. ### Wrong behavior Instead of Upload more documents. Use Change the responsible instruction and rerun the conversation branch. ### Wrong field Instead of Repair reports manually. Use Clarify the output description and confirmation step. ### Wrong action Instead of Hide the tool error. Use Inspect invocation arguments and result, then improve validation and fallback. ## Production readiness checklist Complete this review with a product owner and, when applicable, legal, compliance, security, or operations stakeholders. **0 of 12 complete**Loading saved progress… **01**The agent has one primary job, audience, definition of success, and explicit out-of-scope behavior.**02**All six builder areas were reviewed; the generated draft was not accepted without human review.**03**Instructions cover identity, language, known context, sequence, knowledge and tools, guardrails, recovery, and closing.**04**Every input placeholder has a safe default and every output has a precise expected shape.**05**Only current, approved PDF or website knowledge is attached to the agent.**06**Every configured tool has a tested success result, failure result, authorization boundary, and caller-facing fallback.**07**The voice is understandable in every configured language; any clone has documented permission and approved use.**08**Silence, interruptions, turn-ending, transfer, end phrases, maximum duration, and post-call reporting are tested.**09**The complete scenario suite passes in Draft using both conversational and channel-specific testing where needed.**10**The selected Live agent is connected to the intended phone number, direction, or controlled outbound request.**11**Consent, disclosure, do-not-contact, data minimization, recording, and retention requirements are handled for the deployment context.**12**A named owner is monitoring the first calls and can pause traffic or roll back immediately.