Strongly Typed Execution - Compose.Market

Manowar lets the model reason, but keeps execution contracts typed. The clearest example is the Manowar Agent Loop (mal) harness.

Typed Surfaces

Surface	Type source	Runtime behavior
Tools	JSON Schema or Zod	Converted into OpenAI-style function tools.
MAL plans	`harness/swarm/types.ts` plus interpreter validation	Parsed from YAML/JSON and dispatched by closed `op` enum.
Subagents	`SubAgentSpec`	Requires deployed `agentWallet`, explicit model, budgets, and depth.
Memory loop	`AgentMemoryScope` and workflow response types	Enforces scope and step names.
Telemetry	`UsageRecord`, `FeedbackRecord`, `RunMetadataRecord`	Normalizes provider usage into runtime records.

MAL Ops

Op	Purpose
`task`	Spawn a focused registered subagent.
`delegate`	Call a registered deployed agent.
`fanout`	Run branches with controlled concurrency.
`tool`	Call a registered tool directly.
`if`	Branch on a safe mini expression.
`loop`	Repeat a bounded block.
`scratch`	Read/write private run state.
`synthesize`	Combine saved values with an explicit model.
`stop`	Return final output.
`ask_user`	Request host-provided input.

Example

description: "Research, draft, and synthesize"
budget:
  maxDepth: 3
  maxWallMs: 300000
steps:
  - op: delegate
    saveAs: research
    agentWallet: "0x1111111111111111111111111111111111111111"
    prompt: "Collect source-backed notes for the topic."

  - op: synthesize
    saveAs: answer
    from: ["research"]
    model: "gpt-5.5"
    instruction: "Write a concise answer using only the saved notes."

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

The interpreter validates the plan, resolves {{step.path}} references, executes each step, and records step results. The model does not parse an English instruction manual on every step.