Technical Deep Dive

Platform
Capabilities

A deep dive into the autonomous agent architecture powering every Forged AI platform. 250+ actions, enterprise guardrails, and multi-tier memory — applicable to any industry.

250+
Autonomous Actions
6
Specialized Agents per Platform
4-Tier
Memory System
Real-time
Monitoring
Execution Engine

How It Works — The Agent Loop

A ReAct-pattern execution engine that intercepts every tool call for logging, guardrail checks, cost tracking, and trace recording.

Core Execution Pattern

1
User sends natural language request
2
Agent runtime selects tools from action catalog
3
Each call passes through the Guardrail Pipeline
4
Action executes against connected systems via API
5
Result recorded with full audit trail
6
Agent reflects and continues or returns answer

Fail-Closed by Design

Every action is intercepted. Any guardrail rejection blocks execution and returns an error to the agent — no silent failures, no bypasses.

architecture.txt
   Natural Language (User)
           │
           ▼
   ┌───────────────────┐
   │  Agent Engine      │  ReAct Pattern
   │  6 Specialized     │  Guardrail Pipeline
   │  Agents            │  4-Tier Memory
   │                    │  250+ Actions
   └────────┬──────────┘
            │
            │ REST API (create-from-steps ⚡)
            ▼
   ┌───────────────────┐
   │  Platform API      │  100+ Endpoints
   │  System Adapters   │  Full Audit Trail
   │  Server Validation │  AI describes WHAT
   │                    │  → Server builds HOW
   └────────┬──────────┘
            │
            ▼
   ┌───────────────────┐
   │  Connected Systems │  Databases, APIs
   │  Your Data         │  Services
   │                    │  Real workflows
   │                    │  Real records
   └───────────────────┘
⚡ In Action

Natural Language → Workflow

The agent describes intent in compact JSON. The server builds everything — validated, ordered, and audited.

The Design Principle

“AI Describes WHAT, Server Handles HOW.”

The LLM doesn't generate XML or code — it picks activity IDs from a catalog and describes intent in ~150 tokens. The server creates all records directly, with full validation.

Why This Works:

LLM output stays within token budget (150 vs 2000+)
Server validates ordering and business rules
No code generation, no truncation, no hallucinated success
Same database records as manual UI creation
Full audit trail on every operation
Test mode available for dry-run validation

Token Budget Comparison

Traditional Code Generation2000+ tokens (truncated!)
Step-by-Step API Calls600-1200 tokens
Create From Steps ⚡~150 tokens
create-from-steps.json
// User: "Create an onboarding workflow for new clients"
POST /api/v1/workflow-builder/create-from-steps
{
"name": "Client Onboarding",
"code": "CLT-ONBOARD",
"steps": [
{ "activityId": 76, "name": "Start" },
{ "activityId": 85, "name": "Intake Review" },
{ "activityId": 91, "name": "Document Collection" },
{ "activityId": 95, "name": "Verification" },
{ "activityId": 54, "name": "Complete" }
]
}
✅ 201 Created
{ "templateId": 1045, "levels": 5, "states": 5 }
⏱ 2.1s · 148 tokens · $0.001

The Key Insight

The AI doesn't need to know your system's internals. It picks activity IDs from a catalog and names the steps. The server handles all database record creation — validated, ordered, and audited.

Action Catalog

250+ Actions Across 12 Domains

Every action is self-describing, attribute-registered, and executable through natural language. Here's what the platform covers.

Workflow Builder

46 actions

  • Templates, automation steps, approvals
  • Routing and sequencing logic
  • Create-from-steps (single API call)

Data Processing

22 actions

  • Import, transform, validate, export
  • Schema mapping and conversion
  • Batch operations and ETL

Entity Management

18 actions

  • CRUD operations, fields, relations
  • History tracking and attachments
  • Type and schema management

Task Management

17 actions

  • Create, assign, prioritize
  • Track completion and deadlines
  • Reopen, reassign, escalate

Document Processing

15 actions

  • Upload, analyze, extract
  • Categorize and tag content
  • Version tracking and metadata

Approval Workflows

12 actions

  • Queue, approve, reject
  • Escalate and delegate
  • Full decision audit trail

Process Automation

11 actions

  • Start, advance, pause, resume
  • Monitor execution state
  • Instance lifecycle management

Search & Discovery

10 actions

  • Full-text and filtered search
  • Faceted navigation
  • Saved searches and alerts

Reports & Analytics

10 actions

  • Generate, schedule, export
  • Custom parameters and filters
  • Automated distribution

Contact Management

9 actions

  • Profiles, history, preferences
  • Segmentation and tagging
  • Communication tracking

Agent Orchestration

9 actions

  • Multi-agent pipelines
  • Collaboration and handoffs
  • Automated multi-step execution

Integration Hub

8 actions

  • External APIs and webhooks
  • Data sync and transforms
  • Connector management
Security Layer

Guardrail Pipeline

Every action passes through a multi-stage guardrail pipeline before execution. Fail-closed: any rejection blocks the action.

Permission Guardrail

Checks each action against the agent's allowed and denied lists. Supports wildcards and role-based overrides.

Cost Guardrail

Tracks cumulative token cost per session. Blocks execution if the configured budget threshold is exceeded.

Data Protection Guardrail

Scans tool arguments for sensitive data patterns before execution. Prevents accidental data exposure.

Approval Guardrail

High-risk actions require explicit human approval. Pauses execution until authorized.

Memory Architecture

4-Tier Memory System

Context-aware agents that remember what matters and forget what doesn't.

Working Memory

Current session

Chat history, tool state, user context. Cleared per session.

In-Memory

Episodic Memory

Past interactions

Timestamped interaction records indexed by time and outcome.

Database

Semantic Memory

Domain knowledge

Concept and fact store with optional embedding vectors for retrieval.

Database + Embeddings

Procedural Memory

Learned patterns

Successful action sequences linked to intent patterns for reuse.

Database
LLM Routing

Multi-Model Intelligence

Hierarchical model routing with automatic failover and cost optimization. Configurable per deployment.

RoleModelCost ProfileOptimized ForStatus
PrimaryFast Model (configurable)Low costHigh throughputActive
FallbackReasoning Model (configurable)Higher costComplex tasksStandby
SpecializedDomain-tuned / Embeddings (configurable)VariesRetrieval & classificationOptional

Models are configurable per deployment. The routing layer automatically selects the optimal model based on task complexity, cost budget, and availability. We optimize for cost and speed.

Enterprise Ready

Built for Production

Every feature designed for enterprise deployments with full observability and compliance.

Full Audit Trail

Every action, every decision, every result — logged with timestamps, user context, and execution metadata.

Execution Traces

Detailed step-by-step traces of agent reasoning, tool selection, and result processing.

Guardrail Pipeline

Four-stage security pipeline: permissions, cost limits, data protection, and human approval.

RBAC

Role-based access control with granular action-level permissions and agent-scoped restrictions.

Data Protection

Sensitive data scanning, pattern detection, and automatic redaction before execution.

Cost Tracking

Per-session and per-agent token cost tracking with configurable budget thresholds.

Version Control

Agent configurations, prompts, and action catalogs versioned and auditable.

Multi-LLM Routing

Automatic model selection with fallback chains, cost optimization, and provider redundancy.

Technology

Technology Stack

Modern, proven technologies for reliable production deployments.

Backend

Python
REST API
WebSocket

Frontend

TypeScript
Next.js
React

Security

JWT
RBAC
Encryption

Infrastructure

Containerized
CI/CD
Monitoring