Multi-agent Patterns

Orvanta provides first-class support for multi-agent automation — coordinating multiple AI agents as part of a larger workflow, with human oversight baked in via ConsentGate.

Core concepts

Agent

An agent in Orvanta is a persistent, stateful process that can receive tasks, reason about them, and take actions using tools. Agents are defined as scripts with a designated runtime (typically Python or TypeScript) and are long-running rather than one-shot.

Orchestrator

An orchestrator is a specialised workflow (or agent) that dispatches tasks to one or more worker agents, collects their results, and decides what to do next. Orvanta’s workflow engine acts as the orchestrator layer.

Agent pool

An agent pool is a named group of interchangeable agents. The orchestrator dispatches tasks to the pool and Orvanta schedules them to available agents.

Common patterns

Sequential handoff

One agent hands off its output to the next agent in the chain:

Orchestrator → Agent A (research) → Agent B (summarise) → Agent C (publish)

Best for: pipelines where each step depends on the previous result.

Parallel fan-out

The orchestrator dispatches the same task to multiple agents simultaneously and waits for all results:

Orchestrator ──→ Agent 1
              ──→ Agent 2
              ──→ Agent 3
                → Aggregator

Best for: map-reduce patterns, ensemble models, parallel data processing.

Consensus gate

Multiple agents independently evaluate a proposal; the orchestrator only proceeds if a quorum agrees:

Orchestrator → [Agent 1, Agent 2, Agent 3] → Consensus check → Continue / Reject

Best for: high-stakes decisions where a single agent’s judgement is insufficient.

Human-in-the-loop (ConsentGate)

Any agent output can be routed through a ConsentGate before the workflow proceeds. This is the recommended pattern for agents that take irreversible real-world actions.

Agent → Draft action → ConsentGate (human review) → Execute action

See ConsentGate for configuration details.

Agent communication

Agents communicate through the run context — a shared key-value store that all nodes in a run can read and write. This keeps communication explicit, auditable, and easy to debug.

For long-running agents that need asynchronous communication, Orvanta provides a lightweight message bus (coming soon in general availability).

ConsentGate — Human approval for agent actions.
Workflows — How orchestration is modelled as a workflow.
Runs — Inspecting multi-agent run traces.