Tracks economic state for a LiveBench agent including: - Balance (cash + trading portfolio value) - Token costs separated by channel (LLM

Manages work tasks from gdpval dataset: - Loads tasks from parquet file - Selects daily tasks (random or deterministic) - Tracks task

ClawWork economic tracking configuration. Loaded from ``agents.clawwork`` in ``~/.nanobot/config.json``.

Evaluates submitted work artifacts and determines payment. Uses LLM-based evaluation with category-specific rubrics: - Loads occupation-

LLM-based evaluator that uses category-specific meta-prompts to evaluate agent work artifacts with a 0.0-1.0 score.

Normalized execution result across sandbox providers.

BoxLite backend implementation.

AgentLoop with ClawWork economic tracking and tools.

Mutable state shared across all ClawWork tools within a session.

Token pricing (per 1M tokens).

Create a work artifact file in the sandbox directory.

Choose daily activity: work or learn.

E2B backend implementation.

Return the agent's current economic status.

Learn something new and add it to the knowledge base.

LiveAgent - AI agent for economic survival simulation Core functionality: 1. Economic tracking (balance, token costs, income) 2. Dai

Persistent logger for LiveBench agents

Read a file and return its content.

Submit completed work for evaluation and payment.

Classifies task instructions into occupations and estimates value.

Transparent wrapper that tracks token costs via EconomicTracker.

State for wrap-up workflow

LangGraph-based workflow for artifact collection and submission when iteration limit is reached without task completion.

Agent status model

LiteLLMProvider subclass that captures OpenRouter's cost field. Overrides _parse_response to add 'cost' (dollars) to usage when the raw litel

Economic metrics model

Learning memory entry

Client to connect to multiple MCP servers and get their tools

Provider backend interface. Subclasses MUST set provider_name.

* Props for the ScreenReaderStatusMessage component

Manages a persistent sandbox session across execute_code calls.

Work task model