Harness A2A

MAL swarms use registered Compose agents. task and delegate do not accept raw model ids. That rule is the whole point of the harness. A MAL subagent is a deployed agent with its own wallet, card, model, memory scope, tools, and creator ownership. It is not a temporary prompt frame around the coordinator’s model.

Validation

Before a task or delegate runs, the interpreter calls ensureRegisteredAgent().

Check	Behavior
Missing wallet	Fails with `UnregisteredAgentError`.
Malformed wallet	Fails before network lookup.
`GET /agent/{wallet}` returns 200	Step can execute.
`GET /agent/{wallet}` returns 404	Fails closed.
Network/timeout/5xx	Fails closed.

Lookups are cached for five minutes.

Task vs Delegate

Op	Use
`task`	Spawn a focused subagent with optional tool bindings, persona override, budget, and isolation.
`delegate`	Ask a registered deployed agent to perform a request with a budget.

Both resolve the target agent’s card model. If you want to call a model directly, use models_call. task is useful when the coordinator needs a bounded specialist run with optional tool bindings, persona override, budget, or isolation. delegate is the simpler A2A call: ask a registered deployed agent to do one thing and return the result. In both cases, the target agent remains the execution unit.

Example

steps:
  - op: delegate
    saveAs: analysis
    agentWallet: "0x1111111111111111111111111111111111111111"
    prompt: "Analyze this request and return risk notes."
    budget:
      maxWallMs: 120000

  - op: stop
    output: "{{analysis}}"

Comparison

System	A2A shape	Manowar shape
OpenAI Agents SDK	Handoffs route to another agent object.	`task` and `delegate` route to deployed wallet-addressed agents.
AutoGen	Agents exchange messages inside a team/runtime.	MAL subagents are registered Compose agents with card models and memory scope.
LangGraph	Handoffs are graph edges or commands in app code.	MAL plan ops are interpreted by the hosted runtime.

Why This Matters

Prompt-only subagents are cheap to create but hard to account for. They borrow the parent prompt, parent context window, parent credentials, and parent billing surface. Manowar uses a stricter primitive. The coordinator can compose other agents without dissolving them into its own context. That keeps tokens smaller, permissions clearer, and economic attribution attached to the agent that did the work.

Overview

Memory

Harness

Tools

Validation

Task vs Delegate

Example

Comparison

Why This Matters

​Validation

​Task vs Delegate

​Example

​Comparison

​Why This Matters

​Related

Validation

Task vs Delegate

Example

Comparison

Why This Matters

Related