Technical Documentation

An autonomous entity capable of receiving natural language instructions, decomposing them into executable plans, and invoking digital tools to achieve the goal.

Navigate to the Dashboard > 'Create Agent'. Specify a system prompt framing the persona, define the objective, and map the required tools. The engine compiles this into a sandbox.

Input -> Intent Extraction -> DAG Planning -> Tool Orchestration -> State Persistance -> Human Verification -> Execution.

Agent profiles and schema definitions are instantiated.

The LCM reasoning layer generates a Directed Acyclic Graph based on the prompt constraints.

The executor traverses the graph, passing tool outputs as inputs down the chain.

Success/fail vectors and JSON state dumps are saved to the persistent store.

The agent halts when the terminal node in the DAG is reached or a critical fault occurs.

System prompts are hard-coded to enforce output wrapped within strictly structured JSON schemas, preventing hallucinated formatting.

Provide agents with specific IDs instead of vague names. If using a custom Python script, alias it clearly in the prompt context.

The system assigns urgency weights (0.0 to 1.0) based on contextual cues within the input data stream.

Standard REST POST. Requires an authorization header and a JSON payload containing the prompt and agent configuration.

Returns an execution trace ID immediately, switching to a polling or WebSocket streaming model for telemetry updates.

Tools expect standardized inputs and must return deterministic pass/fail flags alongside standard string/JSON output.

Uses dense vector embeddings to classify incoming threads into standardized categories (e.g., inquiry, complaint, spam).

Configurable thresholds. If a calculated confidence score falls below 0.85, the action halts and flags for human oversight.

Webhooks listen to inbox streams, triggering the orchestration loop automatically upon new deliveries.