Search docs...
DocsDevelopmentAI Agents & Runtimes

AI Agents & Runtimes

AI agent system, runtime execution flow, and built-in agent personalities in LeadForge OS.

2 min read•Edit on GitHub

The AI agent runtime manages structured LLM reasoning loops. It dynamically coordinates local and cloud providers with system tools to qualify prospects, research companies, and orchestrate cold email workflows.

Agent Architecture

The diagram below outlines the reasoning cycle and tool dispatch path of an active agent session:

Rendering Outline

Context Builder

The context builder (context-builder.ts) compiles prompts by merging user templates, company profiles, and scrapers' raw text. It limits context window overflow by truncating long web-page inputs and HTML trees before tokens are sent to the provider.

Tool Executor

The tool executor (tool-executor.ts) maps LLM-requested tool parameters into execution commands, triggering system crawlers or updating database rows.

Response Assembler

The response assembler (response-assembler.ts) parses the output payload of the LLM against target Zod validation schemas. This ensures the UI receives valid data formats.

Built-In Agents

LeadForge OS provides two specialized agent implementations:

Research Agent

  • Objective: Discover and qualify local business prospects.
  • System Prompt: Enforces search criteria, evaluates company website indicators, and calculates opportunity scores.
  • Integrated Tools: search_local_businesses, crawl_company_website, and search_linkedin_profiles.
  • Output Schema: Returns structured company profiles, key personnel titles, and contact targets.

Campaign Assistant

  • Objective: Compose and schedule email sequences.
  • System Prompt: Generates outreach templates customized with company insights and value propositions.
  • Integrated Tools: send_outreach_email and poll_inbox_replies.
  • Output Schema: Returns dynamic outreach draft templates and campaign scheduling timelines.

Constraints and Trade-Offs

  • No Autonomous Loops: Agents do not run without supervision. Every execution requires scheduler dispatch to prevent runaway token costs.
  • Resource Dependency: Running local models (like Llama via Ollama) requires sufficient GPU memory (VRAM). Low VRAM configurations degrade crawler performance.
PREVIOUSRelease & VersioningNEXT Troubleshooting