Native Runtime

The native runtime is the restored LLLM agent system inside the protocol-first package. Use it when prompt rendering, parser repair, dialog lineage, tool-call interrupts, invoker traces, or agent sessions are part of the artifact you need to understand.

The protocol boundary stays small: public packages and services expose typed tactics. Native keeps the transparent machinery behind that boundary.

from lllm.runtimes.native import Dialog, Prompt, Role

system = Prompt(path="agent/system", prompt="You are a {style} assistant.")

dialog = Dialog(owner="planner")
dialog.put_prompt(system, prompt_args={"style": "careful"}, role=Role.SYSTEM)
dialog.put_text("Draft the next checkpoint.", name="operator")

retry = dialog.fork(last_n=1, first_k=1)

Restored Shape

lllm.runtimes.native
  core/        Agent, Dialog, Prompt, Runtime, Tactic, resources, config
  native/      NativeTactic and agent-backed tactic helpers
  invokers/    provider invocation boundary, including LiteLLM
  proxies/     API proxy tools and tactic-backed proxy endpoints
  sandbox/     optional Jupyter execution support
  tools/       optional computer-use helpers

What Native Owns

Area	Page
Prompts, tools, MCP declarations, dialogs, messages	Prompt And Dialog
Tagged parsers, repair handlers, function tools	Parsers And Tools
Agent call loop, sessions, context management	Agent Loop
Resource registry, package discovery, provider invokers	Registry And Invokers
Native tactic sessions and tactic-backed tools	Native Tactics
Adapter boundary into the protocol	Protocol Adapter
Proxies, skills, Jupyter, computer-use helpers	Proxies, Skills, And Sandbox
What was preserved, adapted, or kept out of the public protocol	Boundary And Verification

Architecture

flowchart TD C["Caller"] --> B["Tactic boundary"] B --> A["NativeTacticAdapter"] A --> N["native Tactic or native object"] N --> R["Runtime / Registry"] N --> G["Agent group"] G --> D["Dialog"] G --> P["Prompt"] P --> F["Function / MCP tools"] G --> I["Invoker"] I --> M["Model provider"] F --> X["Proxy / Python / tactic tool"]

Importing lllm does not import provider integrations. LiteLLM, Jupyter, OpenAI computer-use helpers, Playwright, and proxy-specific packages remain optional.

When To Use Native

Use native when you need the inside of the agent turn:

• prompt templates with parsers, repair prompts, handlers, or tool contracts;
• dialog branching, replay, lineage, dataset generation, and debugging;
• function tools, MCP tool declarations, tactic-backed tools, and proxy tools;
• multi-agent tactics with per-agent sessions and cost accounting;
• provider invocation experiments through LiteLLM.

Use Pydantic AI or plain Python when the implementation does not need native prompt/dialog/session machinery.