Skip to content
Docs

AI Agents & LLM Orchestration

Interplay treats AI as a first-class node type rather than a bolt-on. A flow can call any supported model provider, give an agent a set of tools, ground it in a vector store, pause for human approval, and hand the result to the rest of the business process — all on the same canvas.

Which model providers can Interplay call?

Section titled “Which model providers can Interplay call?”

Interplay’s GenAI palette includes nodes for major commercial model providers and for private, self-hosted endpoints. Because the platform speaks the common OpenAI-compatible wire format in addition to provider-native APIs, models served by Lifeboat, or any other OpenAI-compatible inference server, can be used interchangeably with hosted providers. Cloud AI/ML platforms including AWS SageMaker and Google Vertex AI have dedicated integrations. Providers can be mixed inside a single flow — for example, a cheap local model for classification and a frontier model for the final synthesis step.

How do agents work?

Section titled “How do agents work?”

An agent node is configured with instructions, a model, and a set of tools it is allowed to call. During a run the agent decides which tools to invoke, in what order, and when it is finished; the flow captures each step. Agents can:

  • Call tools built from any node in the palette, so an agent can query a database, hit an internal API, read a file, or trigger another flow.
  • Connect to MCP servers (Model Context Protocol) to consume externally published tool sets, with connection testing and a server registry built in.
  • Stream responses back to a caller or a chat widget as they are produced.
  • Run inside conditional and looping logic, so agentic steps compose with deterministic control flow rather than replacing it.
Section titled “Does Interplay support RAG and vector search?”

Yes. Dedicated nodes store embeddings into a vector database and search it by meaning, which is the basis for retrieval-augmented generation. Combined with the file, mail, and document nodes, a typical pattern is: ingest documents → chunk and embed → store → retrieve at query time → ground the model’s answer. Large-dataset handling is built into these nodes so ingestion is not limited to small samples.

Can a human approve an agent’s action?

Section titled “Can a human approve an agent’s action?”

Yes. A user-approval node pauses a flow and waits for a person to approve, reject, or amend before execution continues. This is the standard pattern for agentic workflows that write to systems of record, send external communications, or move money — the agent proposes, a human commits.

Can Interplay fine-tune models?

Section titled “Can Interplay fine-tune models?”

Interplay includes an LLM fine-tuning workflow that detects available CUDA GPUs, runs training jobs, exposes training status and logs, and manages the resulting model artifacts — so teams that need a task-specialized private model can produce one without leaving the platform.

How is this governed?

Section titled “How is this governed?”

Model calls made from Interplay flows can be routed through AgentWatch to add data-loss prevention, guardrails, budget enforcement, and a full audit trail across every provider. Within Interplay itself, credentials for model providers live in the secrets wallet rather than in flow definitions.

Sources and references

Section titled “Sources and references”