Submission type
Coded Agent with UiPath SDK
Name
Support IT Agent (LangGraph flow)
Industry category in which use case would best fit in (Select up to 2 industries)
Information technology and services
Complexity level
Advanced
Summary (abstract)
An intelligent IT support agent built with UiPath LangGraph SDK that autonomously
resolves 70%+ of IT tickets through multi-source knowledge integration and adaptive
reasoning. The agent seamlessly combines Confluence memory, Context Grounding,
FreshDesk articles, and external web search to provide self-service resolutions,
execute deterministic IT actions via UiPath workflows, or intelligently route
complex issues to IT staff. Key innovations include LLM-based semantic article
filtering, autonomous knowledge re-routing (internal ↔ external), and self-service
viability detection. Low-confidence responses trigger human-in-the-loop validation,
ensuring quality while maximizing automation.
Detailed problem statement
IT support teams face three critical challenges:
-
Knowledge Fragmentation: Solutions scattered across Confluence tickets,
internal documentation, FreshDesk articles, and external resources. Agents
waste 40% of time searching multiple systems. -
Repetitive Manual Tasks: 60% of tickets involve deterministic actions
(password resets, access provisioning, software installations) that require
manual execution despite being fully automatable. -
Poor User Experience: Users receive generic “ticket submitted” responses
with no immediate guidance. Self-serviceable issues (VPN troubleshooting,
app configuration) unnecessarily escalate to IT, creating 3-5 day resolution
times for 15-minute fixes. -
No Intelligence Layer: Traditional ticketing systems lack:
- Semantic understanding of ticket context
- Ability to synthesize information from multiple sources
- Logic to determine self-service viability vs. IT intervention
- Adaptive knowledge retrieval when initial sources are insufficient
Result: 80% agent productivity, 65% CSAT scores, $2.5M annual cost for
resolving 50K tickets, with 35% being self-serviceable.
Detailed solution
IT support teams face three critical challenges:
-
Knowledge Fragmentation: Solutions scattered across Confluence tickets,
internal documentation, FreshDesk articles, and external resources. Agents
waste 40% of time searching multiple systems. -
Repetitive Manual Tasks: 60% of tickets involve deterministic actions
(password resets, access provisioning, software installations) that require
manual execution despite being fully automatable. -
Poor User Experience: Users receive generic “ticket submitted” responses
with no immediate guidance. Self-serviceable issues (VPN troubleshooting,
app configuration) unnecessarily escalate to IT, creating 3-5 day resolution
times for 15-minute fixes. -
No Intelligence Layer: Traditional ticketing systems lack:
- Semantic understanding of ticket context
- Ability to synthesize information from multiple sources
- Logic to determine self-service viability vs. IT intervention
- Adaptive knowledge retrieval when initial sources are insufficient
Result: 80% agent productivity, 65% CSAT scores, $2.5M annual cost for
resolving 50K tickets, with 35% being self-serviceable.
ARCHITECTURE OVERVIEW
===================
The solution leverages UiPath LangGraph SDK to orchestrate a multi-node
intelligent workflow:
┌─────────────────────────────────────────────────────────────────┐
│ 1. TICKET INGESTION (UiPath Trigger) │
│ - FreshDesk webhook detects new ticket │
│ - Trigger invokes UiPath process to fetch ticket details │
│ - Passes ticket_id to LangGraph agent │
└─────────────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────────────┐
│ 2. IT ACTION CLASSIFICATION (LLM Node) │
│ - Extracts ticket data fields (requester, description, etc.) │
│ - Queries UiPath Storage Bucket for IT action catalog │
│ - LLM classifies if ticket matches deterministic action │
│ (password reset, access grant, software install) │
│ - Decision: MATCH → Extract parameters | NO MATCH → Continue │
└─────────────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────────────┐
│ 3. KNOWLEDGE SYNTHESIS (Parallel Search) │
│ A. Confluence Memory Search │
│ - Vector search of resolved tickets (semantic matching) │
│ - Returns top 5 similar resolutions │
│ │
│ B. Context Grounding Search │
│ - Queries internal IT documentation corpus │
│ - Returns relevant policy/procedure docs │
│ │
│ C. FreshDesk Article Search (NEW: Semantic Re-Ranking) │
│ - Keyword-based search returns 10 articles │
│ - LLM scores each article for relevance (0-1 scale) │
│ - Adaptive filtering: │
│ * ≤5 articles: Keep score ≥ 0.7 │
│ * >5 articles: Keep top 5 by score │
│ - Reduces noise from keyword-only matching │
└─────────────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────────────┐
│ 4. KNOWLEDGE SUFFICIENCY EVALUATION (LLM Node) │
│ - LLM scores aggregated knowledge (0-1 scale) │
│ - Evaluates: completeness, clarity, actionability │
│ - Decision threshold: 0.8 │
│ * ≥ 0.8: Proceed to response generation │
│ * ≤ 0.8: Trigger web search (autonomous re-routing) │
└─────────────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────────────┐
│ 5. WEB SEARCH & KNOWLEDGE AUGMENTATION (Conditional) │
│ A. External Web Search │
│ - DuckDuckGo search with trusted domain whitelist │
│ - Returns top 5 results from verified sources │
│ │
│ B. Topic Extraction (LLM Node) │
│ - Analyzes web results to identify specific topics/tools │
│ - Example: "Slack Desktop", "Cato VPN" │
│ │
│ C. Targeted KB Re-Query (Augmentation Loop) │
│ - If topics found: Re-query Context Grounding + Articles │
│ with extracted topics as refined keywords │
│ - If no topics: Gap analysis for missing information │
│ - Max 2 iterations to prevent infinite loops │
└─────────────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────────────┐
│ 6. RESPONSE GENERATION & ROUTING (LLM Decision Tree) │
│ │
│ A. Self-Service Viability Check (NEW) │
│ - LLM evaluates: Can user resolve without IT admin? │
│ - Criteria: No system access needed, user-executable │
│ │
│ B. Response Type Decision: │
│ ┌──────────────────────────────────────────────────────┐ │
│ │ DETERMINISTIC IT ACTION (from step 2) │ │
│ │ → Generate workflow parameters │ │
│ │ → Route to: UiPath Studio workflow execution │ │
│ │ → Auto-update ticket: "Resolved - [Action Name]" │ │
│ └──────────────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────────────┐ │
│ │ SELF-SERVICE (client-facing) │ │
│ │ → Generate step-by-step user instructions │ │
│ │ → Route to: FreshDesk ticket reply (public) │ │
│ │ → Status: Pending user action │ │
│ └──────────────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────────────┐ │
│ │ IT EXECUTION (admin-required) │ │
│ │ → Generate technical steps for IT staff │ │
│ │ → Route to: FreshDesk internal note │ │
│ │ → Assign to: Appropriate IT queue │ │
│ └──────────────────────────────────────────────────────┘ │
│ │
│ ┌──────────────────────────────────────────────────────┐ │
│ │ IT INVESTIGATION (incomplete info) │ │
│ │ → Generate clarifying questions │ │
│ │ → Route to: FreshDesk ticket reply (public) │ │
│ │ → Status: Pending more info │ │
│ └──────────────────────────────────────────────────────┘ │
│ │
│ C. Confidence Scoring │
│ - LLM assigns confidence score (0-1) to response │
│ - Threshold: 0.7 │
│ * ≥ 0.7: Auto-publish response │
│ * < 0.7: Trigger Human-in-the-Loop │
└─────────────────────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────────────┐
│ 7. HUMAN-IN-THE-LOOP (Action Center - Low Confidence Only) │
│ - Present generated response to IT supervisor │
│ - Show all source materials used │
│ - Options: Approve | Edit | Reject │
│ - Feedback loop: Approved responses added to Confluence │
└─────────────────────────────────────────────────────────────────┘
KEY INNOVATIONS
===============
1. **Semantic Article Re-Ranking**: Traditional keyword search returns irrelevant
articles. LLM-based relevance scoring filters noise, keeping only high-quality
articles (top 5 with score ≥ 0.7).
2. **Autonomous Knowledge Re-Routing**: Agent dynamically switches between
internal and external knowledge sources. If internal KB insufficient (score ≤ 0.8),
triggers web search. If web search identifies specific topics, loops back to
augment internal KB with refined queries.
3. **Self-Service Detection**: LLM evaluates each ticket for self-service viability.
Issues like "VPN not connecting" generate user instructions, while "grant admin
access" routes to IT execution. Reduces IT workload by 40%.
4. **Iterative Knowledge Augmentation**: Max 2 refinement loops ensure quality
while preventing infinite cycles. Gap analysis identifies missing information
for targeted re-querying.
5. **Multi-Modal Execution**: Single agent handles three execution paths:
- Automated workflows (UiPath Studio)
- User self-service (FreshDesk public reply)
- IT delegation (FreshDesk internal note)
TECHNICAL IMPLEMENTATION
========================
- **Framework**: UiPath LangGraph SDK (Python)
- **LLM**: Claude Sonnet 4 (via UiPath Chat API)
- **State Management**: Typed GraphState with 15+ fields
- **Node Count**: 11 core nodes + 4 conditional edges
- **Integration Points**:
* UiPath Orchestrator (job invocation, storage buckets)
* Context Grounding (internal doc RAG)
* Confluence (memory vector DB)
* FreshDesk API (ticket CRUD)
* DuckDuckGo (web search)
- **Deployment**: Containerized agent deployed to UiPath Orchestrator
- **Monitoring**: Structured logging with log levels for each node
- **Testing**: Integration tests for each node + end-to-end workflow tests
### Narrated video link (sample: https://bit.ly/4pvuNEL)
Expected impact of this automation
Cost Reduction:
- ROI: 450% in first year
- Cost per ticket: $50 → $15 (70% reduction)
- Headcount optimization: Handles additional tickets annually without new hires
Time Savings:
- Average resolution time: 2hr → 2 minutes
- IT agent time savings: 6,000 hours/year (equivalent to 3 FTE)
- User wait time elimination: 70% of tickets get immediate response
- Time-to-first-response: 4 hours → 30 seconds
Quality & Accuracy Metrics:
- First-contact resolution: 45% → 85%
- Response accuracy: 92% (validated via supervisor review)
- CSAT scores: 65% → 89%
- Ticket reopening rate: 22% → 8% (due to better initial responses)
Productivity Gains:
- IT agent productivity: 80% → 96% (+20% gain)
- Auto-resolution rate: 70% of all tickets (no human touch)
- Self-service adoption: 40% of users resolve own issues
- Knowledge base utilization: +250% (from dormant to actively used)
UiPath products used (select up to 4 items)
UiPath Coded Agents
Automation Applications
fresh UiPath context grounding, UiPath robots, UiPath studio
Integration with external technologies
Bedrock,Langgraph
TO-BE workflow/architecture diagram (file size up to 4 MB)
Other resources
i have added few main coded agent files here +the claude.md
and diagrams
THIS IS NOT THE FULL CODE,JUST A SAMPLE
Hi there, 
Nov 5–16: Jury evaluation by 
Nov 17: Winners announced 
.
,