Urjasoft - Software, AI & SaaS Engineering Urjasoft - Software, AI & SaaS Engineering
Prototype ID: LAB-02 Category: Autonomous Agents & LLMs

AI Workflow Copilot for Enterprise Automation

Prototype evaluating schema-validated tool execution loops and audit controls.

An exploratory agent framework testing whether enforcing JSON Schema validation gates and cryptographic approval tokens between LLM decision steps prevents unintended API mutations.

The Architectural Problem

Autonomous AI agents frequently drift or hallucinate unintended API actions when granted unconstrained tool execution permissions in enterprise ERP/CRM environments.

Working Hypothesis

Enforcing JSON Schema validation gates and cryptographic approval tokens between LLM decision steps prevents unintended side effects while retaining automated orchestration speed.

System Topology

Prototype Architecture & Data Pipeline

Multi-stage execution model separating probabilistic reasoning from deterministic state mutations.

1

01. Intent Ingestion

Operator or webhook submits high-level business task or operational event trigger.

Input Interface
2

02. Action Planning

Agent plans tool execution steps with prerequisite dependency checks.

Planning Layer
3

03. Policy Gatekeeper

Deterministic validator checks permissions, rate limits, and payload schemas before any external call.

Security Boundary
4

04. Tool Execution & Audit

Actions execute against internal microservices with HMAC audit signatures and logging.

Safe Dispatch
Data Boundary Rule: No non-deterministic agent loop is permitted to execute writes directly to transactional databases without an intermediate policy gatekeeper.
Implementation Methodology

How We Engineered the Prototype

1. Goal Parsing: Operator intent is parsed into atomic tool actions.
2. Policy Gatekeeper: Deterministic validator checks permissions, rate limits, and payload schemas before any external call.
3. Safe Tool Dispatch: Actions execute against internal APIs with HMAC audit signatures.
4. Human-in-the-Loop Review: Irreversible or financial operations pause automatically for operator authorization.
Current Working Capabilities
  • ✓ [Implemented] Strict JSON Schema validation on model tool-call arguments prior to dispatch
  • ✓ [Implemented] Role-based permission checks before simulating API state changes
  • ✓ [Planned] Directed acyclic graph (DAG) execution across multi-service workflows
Known Limitations & Unsolved Cases
  • × Validation gate introduces 200-400ms latency overhead before tool dispatch
  • × Requires predefined JSON schema contracts for all integrated external tools
Engineering Takeaways & Architectural Findings

“Engineering Hypothesis: Enforcing strict JSON Schema validation boundaries at the dispatch layer prevents parameter drift, but increases end-to-end execution latency across multi-step chains.”

Experiment Parameters
Maturity Status
Prototype
Research Discipline
Autonomous Agents & LLMs
Experiment Identifier
LAB-02
Tested Technologies
Python Laravel Redis OpenAI API
Related Commercial Capability
AI & Automation Solutions

Findings from this research area inform our engineering capabilities in ai & automation solutions.

Explore Related Capability →
Need a custom proof of concept built for your company?
Request Technical Consultation
Additional Explorations

Related Lab Experiments

View All Initiatives →
Engineering Collaboration

Translating research into production software.

Our engineering team can evaluate your technical roadmap, suggest architecture patterns, and partner with you on high-assurance system execution.

Schedule Technical Scoping Call Explore Full Labs Catalog