Agent and Workflow Design

Prev Next

# Agent and Workflow Design

AI agents and workflows are treated as governed, deployable artifacts within the platform.

They coordinate reasoning and tool usage but do not control data access or enforcement.


Tool Invocation via MCP

AI agents invoke tools using Model Context Protocol (MCP).

MCP provides:

  • Typed, inspectable tool definitions
  • Explicit agent-to-dataplane contracts
  • Capability-based exposure

Clarifications:

  • MCP does not store data
  • MCP does not enforce permissions
  • All enforcement remains in the Dataplane

Prompt and Workflow Versioning

Orchestration artifacts are deployable assets:

  • Prompts are versioned
  • Agent definitions are versioned
  • Workflows are versioned

This enables reproducibility, auditability, and rollback.


Environment Promotion

Orchestration follows the same lifecycle model as the platform:

  • Development → Test → Production
  • Promotion is policy-gated
  • Changes are auditable

If behavior differs between environments, it is a defect.


Failure Handling and Observability

Orchestration failures are first-class operational events.

Principles:

  • Failures are explicit
  • Partial execution is observable
  • Tool failures are distinguishable from model failures

Observability captures:

  • Which agent invoked which tool
  • Under which identity
  • Using which pipeline
  • With which outcome

By separating thinking from data access, AI Fabrix enables safe evolution of agents and workflows without reintroducing security, compliance, or audit risk.