What's New for Integrators

This page is for integrator teams (such as ConvAI and other developers who use Wetel as their AI backend) who need a quick view of what the platform can do today, where each capability is documented and what is safe to rely on.

Capability map

Capability	What it gives an integrator	Documentation	Availability
set_variable node	Assign, transform or extract values from run context without a model call. Read channel data, reshape structured output, prepare inputs for later nodes.	Workflow Overview	Production
switch node	Branch the workflow on expressions over run variables (e.g. route by error code, customer tier or structured output field).	Workflow Overview	Production
Structured output (outputSchema)	Force an llm node to return validated JSON matching a schema, so your integration can parse the result programmatically without brittle text extraction.	Workflow Overview	Production
Retry policy	Automatically retry tool, action and webhook nodes on transient failures (TIMEOUT, UNREACHABLE, RATE_LIMITED, UPSTREAM_ERROR) with configurable backoff, idempotency guard and deadline.	Build a Reliable Workflow	Production
Failure routing (success/failure edges)	Route node failures to a recovery path instead of ending the conversation. Access _lastNodeErrorCode and _lastNodeErrorPublic on the failure edge.	Fail Gracefully	Production
Mock side-effects mode	Draft tests and regression replays run tool, action, webhook and pdf_export nodes without making real outbound calls, spending quota or writing files. Validates config and templates exactly as live.	Workflow Testing	Production
Approvals	Pause a workflow until a workspace owner approves or rejects, letting a human gate sensitive actions before the agent continues.	Workflow Overview	Production
Streaming sentences (publishSentences)	Send each finished sentence to the client while the model is still writing, as partial AI_RESPONSE events with isFinal: false. Reduces perceived latency for chat UIs.	Workflow Overview	Production
Artifact generation (generate_artifact node)	Have the model produce documents (markdown, HTML), images, audio (speech, music) and video scripts, stored as downloadable artifacts.	Generate Artifact Node, Artifacts API	Off by default
Group chat mode	Run an agent in a group conversation where it answers only when mentioned or replied to, using one session per group with sdkStart/sdkSendMessage and the Group Brain API.	Answer When Mentioned in Group Chats	Production
Telephony	Voice calls to and from the agent over the phone network (SIP/RTP), with a listen-think-speak loop. Requires a test environment and Wetel staff involvement.	Telephony Quickstart, Telephony API Reference	Staging
Role templates and hire	Provision a ready-made agent per customer from a template with a single templates:hire API key call, idempotently, and bind shared resources.	Sub-Agent Inputs and Outputs	Production

Notes on availability

• Production means the capability is live and in the production schema. It may still have per-workspace settings (such as member-built drafts) that an owner must enable.
• Staging means the capability is available on the staging environment only. Do not build production flows on it.
• Off by default means the capability exists in the production schema but is behind a platform switch that Wetel staff must enable for your environment. While the switch is off, the feature fails immediately (e.g. DISABLED for generate_artifact).

Recommended adoption order for a chat-platform integrator

If you are building a chat product on Wetel (as ConvAI does), adopt capabilities in roughly this order:

1. Failure routing and set_variable/switch. These are free, model-call-free and make every workflow more robust. Wire failure edges on every node that calls out, and use _lastNodeErrorCode to branch. See Fail Gracefully and Build a Reliable Workflow.

2. Structured output. Add outputSchema to llm nodes whose output your code parses. This eliminates brittle regex extraction and gives you a typed contract. See the Workflow Overview for the schema constraints.

3. Mock side-effects mode. Run every draft test and regression replay in mock mode before publishing. This catches template errors, missing variables and broken edges without real outbound calls. See Workflow Testing.

4. Retry policies. Add retry to webhook, tool and action nodes that call external services. Mark non-idempotent calls so the platform does not resend them after a timeout. See Build a Reliable Workflow.

5. Streaming sentences. Enable publishSentences on user-facing llm nodes to reduce perceived latency in your chat UI. Test that your client handles partial events (isFinal: false) and the empty final event correctly.

6. Group chat mode. If your product supports group conversations, follow the mention-only recipe to add agent participation without flooding the group.

7. Artifact generation. Once Wetel has confirmed the switches are on for your environment, integrate document, audio and video generation. Start with markdown (simplest), then audio_speech, then video_script. See the companion recipe for a worked example. Remember that video_script produces a JSON manifest, not an MP4.

8. Telephony. Request access to the staging telephony environment when you are ready to prototype voice. See Telephony Quickstart.

Things not to rely on yet

The following are present in the schema or mentioned in documentation but are not yet safe to build production flows on:

• Artifact generation switches. All generate_artifact artifact types are off by default in production. Do not hard-code artifact flows into customer-facing products until Wetel has confirmed they are switched on for your environment. The node fails with DISABLED while the switch is off.
• Video script as MP4. A video_script artifact is a JSON manifest played by the workspace app. “Download video” records an MP4 or WebM in the viewer’s browser. The platform does not render or serve MP4 files server-side. Do not promise server-rendered video to your customers.
• visibility on artifacts. The visibility field (PRIVATE or TENANT) exists on ArtifactDto but is reserved and not enforced. Every artifact follows the access rules of its session and AI employee. Do not rely on visibility for access control.
• Telephony in production. Voice calls are a staging-only capability available on request. Per-key daily call caps exist in the production code, but the telephony feature itself is not generally available.
• Cost attribution (tenantCostByAgent). The query exists in production but attribution is a platform setting that is off by default. While off, agents is empty and attributionEnabled is false.
• Member-built drafts and releases. The full drafts-and-releases workflow (for agents, workflows, knowledge bases and custom actions) is in production but off by default per workspace. An owner must enable it.

When any of these change, the update will appear in the Changelog.

Further reading

• Workflow Overview — the node/edge model, common fields, failure routing, retries.
• Workflow Testing — mock mode, draft tests, regression replays.
• Artifacts API — reading, listing and downloading artifacts.
• Channel Capabilities — what each channel supports today.
• Changelog — all API-facing changes, newest first.