SkillMesh

• 

Inspiration

AI agents are becoming increasingly capable of reasoning, but reasoning alone does not allow an agent to get real enterprise work done.

In a real organization, the capabilities an agent needs are usually spread across different systems such as customer platforms, ticketing systems, knowledge bases, incident-management systems, internal tools, and external services.

This led us to one simple question:

What if enterprise capabilities could become reusable building blocks that any authorized AI agent could discover and use?

That question inspired SkillMesh

Instead of building another chatbot, we explored a capability layer where agents can discover reusable skills, compose them into workflows, access tools through MCP, use enterprise knowledge, operate within governance boundaries, and maintain visibility into their actions.

What We Built

SkillMesh is a prototype agent capability orchestration platform built around reusable enterprise skills.

The platform brings together:

Reusable Skills*for capabilities such as customer lookup, ticket search, knowledge search, incident detection, and ticket updates MCP Tool Registry for exposing capabilities through standardized tool interfaces Knowledge Layer for enterprise documentation, policies, troubleshooting information, and previous incident context Agent Playground for demonstrating agents discovering and composing capabilities Governance for controlling what agents are allowed to do Human Approval for sensitive actions that require human review Execution Trace for visibility into agent actions and tool usage

The core workflow is:

Agent → Discover Skills → Compose Workflow → MCP Tools → Enterprise Systems → Govern → Execute → Trace

How It Works

We demonstrate SkillMesh using an enterprise investigation scenario.

For example:

Investigate why ABC Corporation's production synchronization stopped.

Instead of giving the agent one hard-coded workflow, SkillMesh demonstrates how the agent can identify the capabilities required for the task.

It can discover capabilities such as:

Customer Lookup → Ticket Search → Knowledge Search → Similar Incident Detection

These capabilities can then be composed into an investigation workflow.

The agent gathers relevant enterprise context, combines the information as evidence, and produces a recommendation.

If the next step involves a sensitive operation, governance rules can require human approval before the action is executed.

The resulting workflow is:

Understand → Discover → Compose → Retrieve → Reason → Approve → Execute → Trace

MCP and Freshworks

Track 2 focuses on reusable skills, MCP integrations, and the Freshworks developer platform.

Our approach is not to replace existing enterprise platforms.

Instead, SkillMesh explores how capabilities from platforms such as Freshworks can become accessible to AI agents through an agent-oriented capability layer.

For example:

Ticket Search Skill

↓

ticket.search MCP Tool

↓

Enterprise System

The prototype demonstrates this integration pattern using simulated Freshworks capability data.

The objective is to demonstrate how an agent could discover and orchestrate existing enterprise capabilities rather than requiring every capability to be rebuilt inside every individual agent.

Why SkillMesh Is Different

A traditional AI application can follow a model such as:

User → Chatbot → Fixed Tools → Response

SkillMesh explores a different model:

Agent → Skill Discovery → Skill Composition → MCP → Enterprise Systems → Governance → Execution Trace

The important shift is from building an agent with a fixed collection of tools to creating a platform where capabilities can be discovered, reused, composed, governed, and observed

This makes the capability itself a reusable building block rather than functionality hidden inside one specific agent.

What We Learned

Building SkillMesh helped us understand that agentic AI is much more than connecting an LLM to a few APIs.

We learned about:

• MCP and tool-based agent interaction
• Reusable agent capabilities
• Agentic workflow orchestration
• Skill discovery and composition
• Enterprise knowledge retrieval
• Human-in-the-loop systems
• Permission and risk-based governance
• Tool execution and observability
• Designing AI systems around existing enterprise platforms

One of our biggest lessons was that agent intelligence and agent capability are different problems.

An agent may be excellent at reasoning, but without reliable tools, enterprise context, permissions, and execution mechanisms, it cannot reliably perform real work.

Challenges We Faced

One of our biggest challenges was deciding what should belong inside an agent and what should become a reusable platform capability.

We also had to carefully define the relationship between skills and MCP tools.

A skill represents a reusable capability from the agent's perspective, while an MCP tool provides an interface through which that capability can be accessed.

Another challenge was designing governance without eliminating agent autonomy.

We wanted the agent to act automatically where appropriate while ensuring that higher-risk actions could require human intervention.

Finally, because this project was developed as a prototype, we had to balance demonstrating the architecture convincingly with being transparent about what is simulated versus what would require a production integration.

Future Vision

SkillMesh is currently a prototype, but the larger vision is an agent capability ecosystem

Organizations could eventually build and register capabilities once and allow multiple authorized agents and workflows to discover and reuse them.

The ecosystem could expand across:

Customer Support • IT • HR • Finance • Security • Knowledge Management • External Services

with stronger authentication, permissions, monitoring, and production integrations.

Our Vision

AI agents should not only be able to reason.

They should be able to access reliable capabilities, use enterprise context, operate within clearly defined boundaries, and perform real work.

The future of enterprise AI is not just about making agents smarter. It is about giving them reliable capabilities to get real work done.

SkillMesh

Capabilities, not just conversations.

Built With

• agentic
• agents
• ai
• application
• context
• css
• developer
• enterprise
• freshworks
• human-in-the-loop
• knowledge
• llm
• mcp
• orchestration
• platform
• protocol)
• react
• retrieval
• skill
• tailwind
• typescript
• vite
• web