10Chapter 10 of 10

Limits & troubleshooting

Every size limit, cap, and known restriction for voice agents, in one place.

Read time~ 5 min

Sections8

UpdatedJuly 27, 2026

Every size limit, cap, and known restriction for voice agents, in one place.

Agents

Number of agents per tenant — no limit.
Account number — 2 to 32 digits, must be unique, can't be changed after creation.
Description — up to 1,000 characters.
Mode (Inbound/Outbound) — fixed at creation, can't be changed later.

Agent name — up to 60 characters.
Business name — up to 80 characters.
Industry — up to 60 characters.
Role — up to 80 characters.
Primary goal — up to 240 characters.
Tone — up to 80 characters.
Business context — up to 800 characters (longer input is trimmed, with a note, rather than rejected).
Operating instructions — up to 1,000 characters.
Off-script policy — up to 400 characters.
Allowed topics / escalation topics — up to 10 entries each, 60 characters per entry.
Total rendered instructions — capped at 2,200 characters across all of the above combined.
Voice speed — adjustable 0.8×–1.4× on OpenAI and Grok agents; not adjustable on Gemini agents.
Model choice — you choose the engine (Gemini, OpenAI, or Grok); the specific model isn't selectable and is kept current automatically.
Language — English only at this stage, in four variants (English, Australia, UK, US). No other languages are selectable.

See Agent knowledge for the full guide. Quick reference:

File uploads — PDF, DOCX, TXT, or Markdown; 2 MB max per file; 200,000 characters read per item; PDFs are read to their first 20 pages only (later pages are silently skipped, so keep reference PDFs short or split them).
Website import — Zendesk and Freshdesk help centres, or any public site with a browsable articles section; up to 200 articles per import pass; running the import again automatically skips articles you've already brought in, so repeated passes work through a large site; no cap on the total number of imported articles.
Pasted text / manual items — same 200,000-character read limit; no cap on the number of manual items.
Q&A pairs — no length limit on the question or answer; used for word-for-word answers, not searched the same way as regular knowledge, so the 200,000-character limit doesn't apply to them.
No total tenant-wide knowledge size cap — the limits above are per file, per item, and per import pass, not a ceiling on your library.

Providers — Google Calendar and Microsoft Outlook/365 only.
Booking rules are not range-validated — slot length, buffer time, minimum notice, and how far ahead bookings are allowed are all set by you per booking step; the portal doesn't currently stop you entering an unusual combination (e.g. a zero-minute slot), so use sensible values.
Microsoft calendar lookup is weaker than Google's for fuzzy "find my booking" matching — if a tenant relies heavily on cancel/reschedule-by-phone, Google Calendar gives a more reliable experience.
A booking agent can't be published if its calendar connection is missing, not active, or no longer writable — reconnect the calendar under Settings → General before republishing.

Contact list size — up to 1,000 rows per campaign upload; at least 1 row required; duplicate phone numbers within a list, or numbers already in the same campaign, are rejected.
Concurrent calls — a hard tenant-wide ceiling of 50 simultaneous outbound calls, further limited by your configured concurrency and, for carrier-managed tenants, a calls-per-second pacing limit.
Per-campaign concurrency — settable 1–50; defaults to the agent's own configured cap.
Retries — 1 to 5 attempts, with a configurable delay of 0–1,440 minutes between attempts. Which outcomes trigger a retry (no answer, busy, failed, etc.) is set on the agent itself, not adjustable per campaign.
Calling hours — an optional daily window (with timezone) restricts when a campaign is allowed to dial.
Consent — launching a real campaign requires confirming you have a lawful basis to contact the list; if the agent records full calls, confirming your recording-disclosure practice is also required.
Do-not-contact — recipients who've opted out are automatically skipped, both when the campaign is created and again at dial time.
Budgets — optional daily and monthly spend caps per agent; a campaign stops dispatching once a hard cap is hit.

No blind transfer — every hand-off to a person is a warm transfer (the agent briefs them before connecting the caller).
No keypad (DTMF) input — agents are voice-only; don't design a flow that expects a caller to press digits.
No sandbox/test mode — testing an agent means publishing it and placing a real call.
Maximum call duration — configurable per agent; a safety timeout ends the call and routes to your fallback destination if it's ever exceeded.
Caller silence timeout — around 25 seconds by default; two silent turns in a row escalate to your fallback destination.
Evidence/key-moment recording — if enabled on a step, each captured clip is capped at 120 seconds.

Recording follows the same policy options as other call-flow objects (off, record, transcribe, or both), set at publish time.
Transcription is done by VoxaLink's own speech-to-text, not a pass-through of the underlying platform's transcription.

Agent won't publish — check the error message: it's almost always one specific missing requirement (a calendar not connected, SMS not enabled for a texting step, no fallback destination configured, or no human-handoff path defined). Publish validation is strict on purpose, so a caller is never left with no way out of the conversation.
Caller hits a long silence while the agent looks something up — this shouldn't happen; agents are designed to say a short "just checking" line before anything that takes a moment. If you hear real dead air, treat it as a bug and report it rather than a configuration issue.
Website import stalls or comes back empty — the source site must be publicly reachable with no sign-in, and pages blocked by robots.txt are skipped by design; try pointing the import at a more specific section of the site (e.g. yoursite.com/kb) rather than the whole domain.
Import "already in your knowledge" — running Find articles again on a site you've already imported is expected to skip what's already there and only bring in new pages; that's not an error.