Skip to content

Workflow Orchestration

Jarvis Registry workflow orchestration lets you compose A2A agents, MCP tools, and skills into multi-step pipelines that complete complex tasks — with conditional logic, parallel execution, and human approval gates built in.

Composing Agents, Skills, and MCP Tools

A workflow step can invoke any resource registered in Jarvis Registry:

  • A2A Agent step: delegate a subtask to a worker agent (AgentCore, Azure AI Foundry, or self-hosted) and use its output as input to the next step
  • MCP Tool step: invoke a specific tool, prompt, or resource from any registered MCP server
  • Skill step: describe the capability needed in natural language — the Skill Gateway resolves the best-matched agent or MCP server at runtime

Steps are chained by passing outputs forward. The gateway handles transport negotiation, auth prerequisites, and ACL enforcement at each hop transparently.

Workflow Logic

Sequential steps are the default — each step runs after the previous one completes and its output is available.

Parallel execution: steps that don't depend on each other can run concurrently. The workflow waits for all parallel branches to complete before merging results into the next step.

Step 1 (fetch data)
  ├── Step 2a (analyze sentiment)   ─┐
  └── Step 2b (extract entities)    ─┤ parallel
                                     ↓
                              Step 3 (summarize)

Conditional branching: route execution to different steps based on the output of a previous step — for example, escalate to a senior agent if a confidence score falls below a threshold, or skip a step if a resource is already cached.

Approval gates: pause a workflow and require explicit human or system approval before proceeding. The workflow holds state until the gate is resolved — approved, rejected, or timed out — then continues or terminates accordingly.

Retry and fallback: steps can be configured with retry limits and fallback targets — if the primary agent or tool fails, the workflow reroutes to an alternative without failing the whole pipeline.

Next Steps