Skip to content

Job Agents

39 domain-specific AI assistants built on a shared framework -- from accountants to trades workers, each with a profession-specific MCP tool surface.

Overview

Job Agents are G6's application layer. While other clusters provide cognitive infrastructure (retrieval, safety, planning), job agents combine those capabilities into domain-specific assistants that understand the terminology, workflows, regulations, and best practices of a particular profession.

Every job agent is built on job_framework, which provides the shared scaffolding: session management, tool registration, domain knowledge loading, conversation history, and MCP server integration. Individual agents extend the framework with profession-specific tools, prompts, and knowledge bases.

The framework enforces a consistent structure: each agent exposes a bounded MCP tool surface covering its domain, uses the same input/output schemas, and follows the same lifecycle (initialize, load domain knowledge, serve tools, persist state). Most agents expose 25 tools; a small number expose more or fewer where the implemented domain surface differs. This consistency means that orchestrators can treat all 39 agents interchangeably at the protocol level while each agent provides deeply specialised capabilities.

Professional review and live integrations

Job agents are domain workflow assistants, not replacements for licensed professionals, accountable operators, or regulated compliance sign-off. High-stakes domains can safety-gate outputs, attach disclaimers, require human review context, or return confidence=0.0 when the requested action should not proceed unattended.

Live web search and live LLM-backed debate are disabled by default. Set G6_JOB_LIVE_SEARCH=1 or G6_JOB_LIVE_LLM=1 only in runtimes where network access, credentials, latency, and model costs are expected.

For engineering work specifically, job_engineer is paid-pilot ready as a planning, review, and reliability aid, but it is not construction, fabrication, deployment, safety, regulatory, or licensed professional-engineering approval. Qualified engineering review remains mandatory for real-world decisions.

For legal work specifically, job_lawyer is MVP/pilot ready as research, drafting, review, checklist, and triage support. It is not a lawyer, law firm, legal-opinion engine, compliance certification system, court-filing authority, or substitute for licensed legal advice. Jurisdiction-specific legal output requires structured professional attestation, and that attestation is structural rather than independent credential verification.

Job Framework

job_framework is the shared component that all 39 agents extend. It provides:

  • Session management -- persistent conversations with state across turns
  • Tool registration -- declarative tool definitions with schema validation
  • Domain knowledge -- structured loading of profession-specific knowledge bases
  • MCP integration -- automatic MCP server generation from tool definitions
  • Lifecycle hooks -- initialize, pre-process, post-process, shutdown

Agent Explorer

Architecture

graph TD
    FW[job_framework] --> CORE[core.AIBlock]
    FW --> MCP[MCP Server]
    FW --> KB[Domain Knowledge]
    FW --> STATE[Session State]

    ACC[job_accountant] --> FW
    LAW[job_lawyer] --> FW
    ENG[job_engineer] --> FW
    MED[job_medical_surgical] --> FW
    IT[job_it] --> FW
    DOT[...34 more agents...] --> FW

Key Patterns

Framework Extension. Each job agent extends job_framework with three customisation points: (1) domain-specific tool definitions, (2) profession-specific prompt templates, and (3) curated knowledge base entries. The framework handles all common concerns (session management, MCP serving, state persistence).

Bounded Tool Surface. Job agents expose compact MCP tool sets, usually 25 tools, with explicit exceptions where the implemented domain surface differs. This forces tool designers to prioritise the most valuable operations for each profession while keeping resource consumption predictable across agents.

Category-Based Discovery. Agents are tagged with categories (Healthcare, Legal & Finance, Engineering & IT, Creative, Operations, Science & Research, Services) enabling orchestrators to select appropriate agents based on task domain without hardcoding agent names.