npi

Provides a standardized action library that abstracts function-calling across multiple LLM providers (OpenAI, Anthropic, etc.) through a unified schema-based registry. Developers define Python functions as actions, which are automatically converted to provider-specific function-calling schemas and routed to the appropriate LLM backend, enabling agents to invoke tools without provider-specific boilerplate.

Exposes a set of pre-built actions for browser automation (navigation, clicking, form filling, screenshot capture, text extraction) that agents can invoke to interact with web pages. These actions are wrapped as callable functions within the action registry, allowing LLM agents to autonomously browse and manipulate web content without direct Selenium/Playwright code.

Enables agents to break down high-level user requests into sequences of discrete actions by leveraging LLM reasoning to plan execution steps. The agent analyzes the user intent, determines which actions from the registry are needed, orders them logically, and executes them sequentially or conditionally based on intermediate results, implementing a form of chain-of-thought planning within the action execution loop.

Maintains conversation history and context across multiple agent-user interactions, allowing agents to reference previous messages, build on prior decisions, and maintain state throughout a session. The agent uses this persistent context to inform action selection and planning, enabling coherent multi-turn workflows where each turn builds on the accumulated conversation history.

Automatically validates action execution results against expected output types and schemas, detects failures or unexpected responses, and implements configurable retry strategies (exponential backoff, circuit breakers) to recover from transient errors. Failed actions are logged with context, and agents can inspect error details to decide whether to retry, skip, or replan the remaining workflow.

Allows developers to define custom actions by decorating Python functions with action metadata (name, description, parameters), which are automatically registered and made available to the agent. The registry is dynamic — new actions can be added at runtime without restarting the agent, and actions can be conditionally enabled/disabled based on agent state or user permissions.

Records detailed execution traces for each agent step, including action invocations, parameters, results, and reasoning decisions. Developers can inspect these traces to understand why an agent made specific choices, debug planning failures, and optimize action sequences. Traces include timing information, error details, and intermediate state snapshots.

Allows agents to execute actions conditionally based on agent state, previous action results, or user-defined predicates. Agents can branch execution paths (if-then-else logic) based on intermediate results, enabling adaptive workflows that respond to changing conditions without requiring explicit replanning. Conditions are evaluated at runtime and can reference action outputs, context variables, and agent state.

Provides templating and formatting capabilities to structure agent outputs according to user-defined schemas or templates. Agents can generate responses in specific formats (JSON, markdown, HTML, plain text) and validate outputs against expected schemas before returning them to users. This enables consistent, structured responses from agents regardless of the underlying LLM's output format.

Collects and aggregates metrics about agent performance, including action execution times, success/failure rates, token usage, and cost estimates. Metrics are tracked per action, per session, and globally, enabling developers to identify bottlenecks, optimize expensive operations, and monitor agent health in production. Metrics can be exported to external monitoring systems.