Agent2Agent — Proof of Concept

This repository contains a proof-of-concept multi-agent system built with Microsoft Semantic Kernel and ASP.NET Core. It demonstrates agent-to-agent (A2A) registration, discovery, inter-agent task delegation, and a Blazor web frontend for user interaction.

Quick architecture summary

• Orchestrator (AppHost) launches and health-checks agents and the Web frontend.
• Agent A (Customer Advocate) — user-facing semantic agent (POST /api/agent/chat).
• Agent B (Registry) — manages agent registration/discovery; A2A endpoints (/tasks, /.well-known/agent.json).
• Agent C (Knowledge Graph) — vector search and factual grounding (Redis + embeddings).
• Agent D (Internet Search) — external search and caching.
• DatasetCreator — ingests CSV/PDF, creates embeddings, populates Redis for Agent C.

Project structure (high level)

• Agent2Agent.AppHost — orchestrator and entry point.
• Agent2Agent.Web — Blazor Server web UI (chat frontend).
• Agent2Agent.AgentA — Customer Advocate agent (POST /api/agent/chat).
• Agent2Agent.AgentB — Registry agent (A2A server).
• Agent2Agent.AgentC — Knowledge Graph agent (vector store + embeddings).
• Agent2Agent.AgentD — Internet Search agent (external search + caching).
• Agent2Agent.ServiceDefaults — shared bootstrap/middleware/extensions.
• DatasetCreator — data ingestion and embedding pipeline.
• Docs/ — architecture and agent role documentation.

Prerequisites

• .NET 9 SDK
• Docker (to run Redis via docker-compose)
• OpenAI API key (or other embedding/chat provider) configured in each agent's appsettings

Getting started (local development)

1. Clone repository
2. Restore and build

dotnet restore
dotnet build

3. Start Redis via Docker Compose (uses docker_compose.yaml)

docker-compose -f docker_compose.yaml up -d

4. (Optional) Ingest sample data into Redis

dotnet run --project DatasetCreator

See DatasetCreator/README.md for options and sample data formats.

5. Launch the orchestrator (starts AppHost, Web, and configured agents)

dotnet run --project Agent2Agent.AppHost

6. Open the Web frontend in your browser (default) https://localhost:5000 If custom URLs are configured in launchSettings.json, use those instead.

How the system works (short)

• The Web frontend sends user queries to AgentA.
• AgentA uses Semantic Kernel chat completion to craft responses and may invoke other agents for factual enrichment.
• AgentB (Registry) provides discovery and delegates tasks to AgentC (KnowledgeGraph) and AgentD (InternetSearch) as needed.
• DatasetCreator populates Redis so AgentC can answer fact-based queries through semantic vector search.

Documentation and further reading

• Docs/architecture.md — architecture overview, diagrams, startup sequence.
• Docs/agents.md — agent roles, endpoints, and sequences.
• Agent-specific READMEs: Agent2Agent.AgentA/README.md, Agent2Agent.AgentB/README.md, etc.
• A2A Spec: https://a2aproject.github.io/A2A/v0.2.5/
• Microsoft Semantic Kernel: https://learn.microsoft.com/en-us/semantic-kernel/

Troubleshooting (common)

• OpenAI errors: verify API key, model id, and network access.
• Redis: ensure docker container is running and connection string matches.
• Inter-agent calls: ensure agents are running and reachable; check logs for plugin/DI registration errors.

Contributing

Contributions welcome. Open issues or PRs and include tests or usage notes where relevant.

License

MIT