2026-11-06 · agents · A2A · MCP

Delegation is not tool invocation

Once an orchestrator agent can talk to other agents, it is tempting to model everything as an agent: test execution, ticket creation, document search. In an agentic testing platform I worked on, we drew a different line — and it made the system easier to reason about and safer to run.

Two different contracts

A2A delegation hands a task to another agent that owns a reasoning workflow. The orchestrator selects a skill, a task manager creates a child conversation and dispatches it, and the specialist returns a result. The child task has its own state and can be resumed.

An MCP tool call asks for a bounded, authorized backend action. Test discovery and test-run creation are domain tool calls; the MCP layer mediates identity, permission checks and typed API operations.

Prefer explicit task delegation when another agent owns a reasoning workflow. Prefer MCP tools when a bounded, authorized backend action is needed.

Why it matters

• Execution is a domain action. A test cloud does not become a sub-agent merely because an AI request started the run. The service stays the source of truth for permissions and results.
• Clear ownership. A specialist agent (test-case generation, quality insights, analysis) is a task owner with its own context; a tool is a function with a schema and an access policy.
• Smaller blast radius. Writes go through a bounded allowlist with approval gates, instead of through an agent that might improvise.

The rule of thumb

Let models interpret intent and synthesize language. Gate permissions, calculations and write operations in code. Delegation buys you focused reasoning; tools buy you guarantees — don't trade one for the other by accident.