Share
Embed
Transcript
41 slides
-
What is AI Agent? Agentic AI? Ref : Techcrunch “No
one knows what the hell an AI agent is“ -
Who am I? Makan Sepehrifar • Staff Engineer @ Proud
Winner of the CTF Challenge @ Code Nomads -
What is an Agent? “Autonomous, Goal-oriented computational system that perform
certain tasks on behalf of a user, role or another program” Agent -
What is an AI Agent? An Agent that uses Artificial
Intelligence (typically a Large Language Model) to achieve its goal. AI Agent Reusable Prompt Agent Large Language Model -
Multi-Agent System Why we needed multiple Agents? • • •
Parallel Execution and Maximum Output Fault Tolerance and Resilience Separation of Concerns and Specialization -
(Multi-) Agentic System Break the Goal into Tasks Execute the
Tasks Aggregate the Results -
Agentic AI vs AI Agents Agentic AI Definition AI System
that uses Agents (and not necessarily AI Agents) to achieve the Goal. -
Agentic AI Shared Context in Natural Language AI Agent Automated
Agent Selection Natural Language Natural Language Summarization Intent Detection Break the Goal into Tasks Execute the Tasks Aggregate the Results -
AI Agent vs Agentic AI Plan a 5-day Amsterdam to
Barcelona trip. Strict budget: €1500 (flights, lodging, Strict Budget : €1500 activities). Generate a detailed itinerary focusing on historical sites and local food. Include the three cheapest flight options and three budget accommodations (hostels/budget hotels). -
Why do we need multiple AI Agents? Homogeneity and Noise
Reduction Context Purity Control, Monitoring and Observability Predictability and Integration Ref : A small number of samples can poison LLMs of any size by Anthropic -
Agentic Architecture Emerging Challenge is How do these Agents (properly)
communicate with each other? How to Scale? Is it possible to develop a manageable distributed agentic system that operates beyond the restrictions of local computing? -
2010 Hype of Microservice Architecture How do Microservices communicate with
each other? Is there a way to abstract the integration? -
Capability Mesh Capability Mesh is a Multi-Agent Architecture where: •
• • • Agents are matched by their Capability (Intent) instead of their name. Unified Integration Protocol decouples Network and Agent Integration. Agent Execution is encapsulated in its Capability. Execution Flow is audited to monitor Agentic AI's Non-Deterministic Nature. -
Here come the Protocols MCP : Model Context Protocol From
AI Agent to any Agent (Tool) A2A : Agent to Agent From AI Agent to AI Agent -
"MCP is revolutionizing AI integration, just as REST transformed the
internet." – Makan Sepehrifar , 2025 -
Again here are the Protocols MCP : Model Context Protocol
• • Standard MCP Registry by Anthropic ◦ Provide search and list of MCP Description RPC 2.0 Built-In tools listing A2A : Agent to Agent • • Capability structured in Agent Cards ◦ https://<agent-host>/.well-known/agent.json Community Project : a2a-registry -
Orchestrator Responsible Entity in an Agentic (AI) framework aiming to
achieve a bigger Goal. Type of Orchestrators: • Key Components: Workflow Managers • Orchestrators with Agentic Capability • AI Orchestrators • • • • • • Workflow Management Integration Capabilities AI Integration Protocol Support Agent Selection (Suggestion) Intent Detection (Bigger Goal Definition) Goal Verification -
Orchestrator Semi-Deterministic Orchestrator: • • Predefined Workflow Predefined Task to
Agent Mapping Non-Deterministic Residue: • • • Input Detection Output Aggregation AI Agent Behaviour -
Orchestrator Non-Deterministic Orchestrator: • Predefined Agent Pool (?) Planning Verification
Execution Non-Deterministic Elements: • • • • • • Input Detection Agent Selection Workflow Organization Goal Verification Output Aggregation AI Agent Behaviour Agent Pool -
Orchestrator Hybrid Orchestrator: • • • Generative AI for Planning
Human-Agent Audit Deterministic Orchestrator Execution Planning Execution Audit Workflow -
Human in the Loop Organizations use “Human Agent” as a
risk mitigation mechanism. Automation Bias (Over-trust) Cognitive Fatigue False Sense of Security Manipulation by AI Explanation Ref : Zombies in the Loop? by Sebastian Krügel -
Human on the Loop Shift from in to on, where
humans are seen as auditors not approvers. Policy as Code Self-Contained Agents • • Deterministic Agents Higher-Level Integration LLM as a Judge • Impact Validation Traceable Human Input -
Last Words Governance & Security AI & Integration Vulnerabilities Correctness
and traceability are a matter of governance, while defense in depth is a matter of security. AI and LLMs solve the problem of discovery and signature unification in integration. They are not interchangeable, but must be combined in your threat modeling. However, this automated discovery introduces vulnerability to context poisoning.