Native Tactics

The restored native Tactic is the callable runtime object used by native workflows. It supports input/output validation, sync and async calls, batch calls, sub-tactic composition, observer hooks, and a TacticCallSession for each invocation.

Agent-Backed Tactics

Subclass NativeTactic when a tactic should build live native agents from config.

from lllm.runtimes.native import NativeTactic


class BriefTactic(NativeTactic):
    name = "brief"
    input_type = str
    output_type = str
    agent_group = ["writer"]

    def call(self, task: str) -> str:
        writer = self.agents["writer"]
        writer.open("draft", prompt_args={"topic": task})
        writer.receive(task)
        return writer.respond().content

Agent Config

config = {
    "invoker": "litellm",
    "global": {
        "model_args": {"temperature": 0.1},
        "max_exception_retry": 3,
        "max_interrupt_steps": 5,
        "max_llm_recall": 0,
    },
    "agent_configs": [
        {
            "name": "writer",
            "model_name": "gpt-4o-mini",
            "system_prompt": "You write concise briefs about {topic}.",
        }
    ],
}

tactic = BriefTactic(config=config)

The parser reads these agent keys:

Key	Meaning
`name`	Agent name. Must match an entry in `agent_group`.
`model_name`	Provider model id passed to the invoker.
`system_prompt`	Inline prompt string.
`system_prompt_path`	Prompt resource path loaded from the native runtime.
`api_type`	`completion` or `response`.
`model_args`	Provider args such as temperature, token limits, headers, or skills payloads.
`max_exception_retry`	Parser/validation repair cap.
`max_interrupt_steps`	Tool-call loop cap.
`max_llm_recall`	Provider/API error recall cap.
`tools`	Tool, tactic, or proxy refs to add to the prompt.
`proxy`	Proxy manager and execution-tool settings.
`context_manager`	Built-in or registered context manager config.
`skills`	Local, URL, wildcard, or provider-hosted skill entries.
`extra_settings`	Native-specific settings reserved for tactic code.

TacticCallSession

NativeTactic creates a fresh tracked agent group per call. The internal _TrackedAgent wrapper records each agent.respond() call into the current TacticCallSession, so session.summary() includes agent-call counts and total cost.

session = tactic("Summarize the native runtime.", return_session=True)

print(session.state)
print(session.summary())
print(session.agent_sessions["writer"][0].delivery.content)

Batch And Async Calls

Native tactics preserve compatibility helpers such as run, arun, bcall, and ccall. Optional stream, astream, events, and aevents surfaces are reported only when a tactic implementation actually provides them.