agent-authoring

# Agent authoring Use this sub-skill when the user wants to write, wrap, migrate, or debug an Agent Lightning agent. ## Route by task | Request | Read/run | | --- | --- | | Create a function-based trainable agent | [references/authoring-workflows.md](references/authoring-workflows.md) | | Fix `@rollout`, `@llm_rollout`, or `@prompt_rollout` errors | [references/troubleshooting.md](references/troubleshooting.md) and [references/api-reference.md](references/api-reference.md) | | Write a class-based `LitAgent` | [references/authoring-workflows.md](references/authoring-workflows.md#class-based-agents) | | Validate one rollout without external services | `python scripts/agent_rollout_smoke.py` | | Debug resources, prompt templates, or returned reward spans | [references/troubleshooting.md](references/troubleshooting.md), then route to [../tracing-and-instrumentation/SKILL.md](../tracing-and-instrumentation/SKILL.md) if spans are involved | ## Key rules - Every agent consumes one task input and some tunable resources. - Function decorators support known signatures, not arbitrary callables. The most common patterns are `def agent(task, prompt_template) -> float` and `def agent(task, llm) -> float`. - Returning a `float` is the simplest final reward path. Returning `None` is valid only when traces and rewards are emitted explicitly. - Class-based agents subclass `LitAgent[T]` and implement `rollout(self, task, resources, rollout)` or async/validation variants. - For local debugging, prefer `OtelTracer`, `InMemoryLightningStore`, and `LitAgentRunner.step` before using multi-process trainer flows. ## Minimal authoring pattern ```python import agentlightning as agl @agl.rollout def my_agent(task: dict, prompt_template: agl.PromptTemplate) -> float: prompt = prompt_template.format(**task) # call tools or an LLM here return 1.0 ``` Run the bundled smoke script when you need an assertion-backed minimal example: ```bash python scripts/agent_rollout_smoke.py ``` ## Boundary This sub-skill covers agent objects and resource injection. For store lifecycle, algorithms, and `Trainer`, use [runner-store-training](../runner-store-training/SKILL.md). For emitters, adapters, and trace analysis, use [tracing-and-instrumentation](../tracing-and-instrumentation/SKILL.md). For CLI services and LLM proxy endpoint checks, use [cli-and-services](../cli-and-services/SKILL.md).