AI Agent Architecture

BASTION deploys 25 specialized AI agents across five functional categories, orchestrated by LangGraph with PostgreSQL-backed state management and strict human-in-the-loop safety controls. The 25 agents comprise 8 LangGraph analysis specialists, 7 COP layer agents, 6 MDMP governance agents, 3 escalation and competition modeling agents, and Ironclaw (Chief of Staff). An additional conceptual layer of JPP staff role templates enables Ironclaw to adapt its behavior per-role context without requiring 102 separate agent processes.

Orchestration

Agents run as LangGraph state machines with PostgreSQL checkpoints for durable execution. Each agent maintains its own checkpoint thread, enabling pause/resume, retry, and audit replay. The checkpoint store persists full graph state including intermediate reasoning steps.

LLM Configuration

Each agent specifies its own LLM provider and model, with system-wide defaults as fallback:

• Provider: Anthropic, OpenAI, or any OpenAI-compatible endpoint
• Model: Configurable per agent (e.g., claude-sonnet-4-20250514 for complex reasoning, lighter models for classification)
• Global defaults: Applied when an agent does not declare a provider/model override

Agent DIDs

Every agent receives a deterministic Decentralized Identifier (DID) derived from its role name and the parent DAO. DIDs provide a stable, cryptographically verifiable identity for message signing, audit attribution, and cross-DAO agent references.

MCP Tool Registry

Agents access external capabilities through the Model Context Protocol (MCP) tool registry:

• Tools are registered with JSON Schema definitions for inputs and outputs
• The registry enforces schema validation on every invocation
• Tools span database queries, blockchain operations, graph traversals, and external API calls

Secure Message Bus

Inter-agent communication runs over a pg-boss job queue with security controls:

Property	Detail
Access control	ABAC-aware; messages filtered by agent role and clearance
TTL	24-hour default; expired messages archived automatically
Dead letter queue	Failed deliveries routed for inspection and retry
Ordering	FIFO within priority bands

Human-in-the-Loop Checkpoints

All consequential agent outputs pass through human review before taking effect. The checkpoint system:

• Pauses execution at configured decision points
• Presents the proposed action with full reasoning chain
• Requires explicit human approval, rejection, or modification
• Logs the decision with the reviewer's identity for audit

Agent Categories

Category	Count	Purpose
LangGraph Analysis	8	Document intelligence, OSINT, threat monitoring, entity resolution, fusion, strategic analysis
COP Layer	7	MIL-STD-2525D symbol generation, entity linking, layer governance, perspective rendering
MDMP Governance	6	Assumption Auditor, Orders Validator, Uncertainty Quantifier, Data Bias Detector, Problem Framing, ROE Compliance
Escalation & Competition	3	Adversary Modeler (MLCOA/MDCOA), Escalation Modeler, Effect Cascader
Chief of Staff	1	Ironclaw — role-adaptive coordinator spanning all JPP staff functions with persistent memory
Total Deployed	25

Note: Earlier versions of this documentation referenced 131 agents, which included 102 JPP staff role templates. These templates are now implemented as role-adaptive context configurations within Ironclaw rather than as 102 separate agent processes. The 25-agent count reflects actually deployed, persistent agent processes.