Strands mapping

loopl keeps the shape of Strands Agents where it reads naturally in Swift, and departs on purpose where the phone is different. If you know Strands, this table is the whole SDK.

Strands (Python)	loopl (Swift)	Notes
Agent(model=…, tools=[…], system_prompt=…)	Agent(model:tools:systemPrompt:config:)	same three arguments, plus an optional AgentConfig
agent("prompt") → AgentResult	try await agent("prompt") → AgentResult	callAsFunction; .text, .iterations, .toolResults, .metrics, .reason
agent.stream_async("prompt")	agent.stream("prompt")	AsyncThrowingStream<AgentEvent, Error>
agent.messages	agent.messages	[Message], Codable, mutable for trimming
@tool def f(x: int) -> str	struct F: Tool { name, description, parameters, call } or ClosureTool(name:description:parameters:) { args in … }	no macros yet; ToolSchema builds the schema
ToolSpec (name, description, inputSchema)	Tool.spec → {type: "function", function: {name, description, parameters}}	the OpenAI/HF chat-template shape
ToolUse (toolUseId, name, input)	ToolCall (id, name, arguments)
ToolResult (toolUseId, status, content)	ToolResult (callId, isError, content, durationMs)
Message + ContentBlock list	enum Message { system, user, assistant(text:toolCalls:), tool(result:callId:name:) }	one case per turn kind — see Messages
Model provider (BedrockModel, OllamaModel, …)	Model protocol (MLXModel, AppleFoundationModel, GGUFModel, FakeModel)	all on-device; actors
model.stream(messages, tool_specs, system_prompt)	model.stream(messages:tools:params:)	the system prompt is messages[0]
StreamEvent (contentBlockDelta, …)	ModelEvent (text, toolCall, metrics)	flattened
callback handler / AgentResult	AgentEvent stream / AgentResult	same two consumption modes
max_iterations (hard stop, default 8–10)	none — Budget.unlimited	opt-in caps: iterations, tool calls, wall time, tokens — see Budget
—	LoopDetector (warns, never stops)	on-device specific
ConversationManager (sliding window)	mutate agent.messages	a built-in window manager is on the roadmap
event_loop_metrics	Metrics on .metrics events and in AgentResult	adds TTFT, tok/s, context used / max
hooks (BeforeToolInvocation, …)	observe AgentEvent	mutating hooks are roadmap
MCP tools	—	MCP servers are network; out of scope for an offline-first SDK
multi-agent (Swarm, Graph)	—	one agent per conversation; compose in your own code

Where loopl deliberately differs

• No iteration cap. The loop is bounded by the user, not by a constant.
• Tools are values with instance properties, not decorated functions. Sendable for free; ClosureTool when you want the one-liner.
• The parser is in the core, not the provider. Small open models speak three tool-call dialects (ToolFamily); ToolCallParser normalises them so the loop is dialect-agnostic and runtimes stay thin.
• Everything is an actor. Models own their weights; the agent owns its transcript; stream is nonisolated.

Same idea, same words

The event loop — stream a turn, run the tool calls, append results, repeat until a turn has no tool call — is Strands' algorithm verbatim. The difference is only where it runs.