Availability and limits

This documentation covers the currently implemented text-agent, agent-versioning, test-chat, and workspace-knowledge surfaces. A capability not described here should not be assumed to be available.

Agent configuration limits#

System prompt: up to 1,000,000 characters.
First message: up to 2,000 characters.
Language choices in Agent Studio: Indonesian (id) and English (en).
Maximum conversation duration: 30 to 3,600 seconds.
Knowledge attachments per agent: up to 50 documents.
max_tokens: -1 for unlimited or an integer from 1 to 32,000. This is an API configuration field, not a dedicated dashboard control.

The dashboard only enables a temperature adjustment when the selected model reports that temperature is adjustable. A listed model ID can still have no active production adapter. See models for the configuration options, active adapters, and exact control behavior.

Knowledge ingestion limits#

Created text and imported URL content: up to 2 MB.
Uploaded source file: up to 20 MB, and extracted text must also stay within the text limit.
Supported file extensions: .txt, .md, .pdf, .docx, .html, and .epub.
A single move or recursive bulk-delete request: 1 to 100 selected entities.

Scanned image-only PDFs are not processed with OCR. URL imports accept public HTTP or HTTPS targets and reject private, loopback, link-local, and other non-public addresses, including after redirects. More diagnosis steps are in knowledge limits and troubleshooting.

Status and safety boundaries#

Knowledge documents move through indexing, ready, or failed. An agent can only publish with referenced documents that are ready. Failed documents remain visible so an admin can diagnose or replace them.

Agent archive is one-way through the current UI and API. Knowledge bulk deletion is permanent and recursive for selected folders. Review agent lifecycle and manage and delete knowledge before those actions.

Agent Testing boundaries#

Agent Testing is available for text Agents on Main. It provides Simulation, Next Reply, and Tool Invocation definitions; workspace folders and sharing; explicit Draft or Live invocation targets; durable asynchronous runs; and result evidence. It does not reproduce voice, phone, Telegram transport, or public-channel delivery.

A definition is limited to 256 KiB and 200 history entries. One Agent configuration can attach 5,000 test references. One invocation expands up to 5,000 entries with 1 to 50 repeats, for at most 250,000 runs. See Usage, safety, and limits before running a large suite.