Skip to main content

LLM Agents

Examples that make real LLM API calls. Each resolves its provider from the environment — set any of the key sets in Choosing a provider and the commands below are identical across OpenAI, Azure OpenAI, Vertex AI, Bedrock, and the rest. Most fall back to a mock provider when no key is set.

graph_with_llm​

StateGraph with real LLM calls inside nodes. This is the most important example for understanding how Chronos combines graph workflows with live LLM reasoning. A classifier node calls the LLM to categorize questions, then conditional edges route to a technical expert (with tools) or a general assistant.

# OpenAI
OPENAI_API_KEY=sk-... go run ./examples/graph_with_llm/

# Azure OpenAI
AZURE_OPENAI_API_KEY=... AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com \
  AZURE_OPENAI_DEPLOYMENT=gpt-4o go run ./examples/graph_with_llm/

# Google Cloud Vertex AI
GOOGLE_CLOUD_PROJECT=my-project GOOGLE_ACCESS_TOKEN=$(gcloud auth print-access-token) \
  go run ./examples/graph_with_llm/

# AWS Bedrock
AWS_REGION=us-east-1 AWS_ACCESS_KEY_ID=... AWS_SECRET_ACCESS_KEY=... \
  go run ./examples/graph_with_llm/

With no provider configured, this example falls back to a local Ollama server at localhost:11434.

Demonstrates:

  • Wiring real LLM providers (OpenAI, Anthropic, Gemini, Azure OpenAI, Vertex AI, Bedrock, Ollama) into graph nodes
  • Conditional routing based on LLM classification output
  • Tool calling within graph nodes
  • Checkpointing with SQLite
  • The YAML equivalent (see examples/yaml-configs/graph-agent.yaml)

See the Building Real-World Agents guide for a detailed walkthrough.

mcp_agent​

Model Context Protocol integration. Connects to an MCP server over stdio, imports every tool it advertises into the agent's registry, and lets the model call them. Uses the official filesystem MCP server.

# Optional live tool server:
npm install -g @modelcontextprotocol/server-filesystem

# Any provider works — the model that calls the MCP tools is picked from env:
OPENAI_API_KEY=sk-... go run ./examples/mcp_agent/

# Azure OpenAI
AZURE_OPENAI_API_KEY=... AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com \
  AZURE_OPENAI_DEPLOYMENT=gpt-4o go run ./examples/mcp_agent/

# Google Cloud Vertex AI
GOOGLE_CLOUD_PROJECT=my-project GOOGLE_ACCESS_TOKEN=$(gcloud auth print-access-token) \
  go run ./examples/mcp_agent/

With no provider set, the tools are still imported and listed — only the model call is skipped.

Demonstrates:

  • AddMCPServer(mcp.ServerConfig{...}) — register an MCP server on the agent
  • ConnectMCP(ctx) — launch servers and import their tools
  • Inspecting imported tools via agent.Tools.List()
  • Graceful degradation when the server binary is absent
  • Any provider (OpenAI, Azure OpenAI, Vertex AI, Bedrock, …) driving the MCP tool calls

See the Model Context Protocol guide for the full workflow.

coding_agent​

Autonomous, Cursor/Aider-style coding agent. Reads, writes, and searches files, runs shell commands (git, build, tests), and uses a vector store for semantic code search (RAG). Runs an autonomous multi-step loop.

# OpenAI
OPENAI_API_KEY=sk-... go run ./examples/coding_agent/

# Azure OpenAI
AZURE_OPENAI_API_KEY=... AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com \
  AZURE_OPENAI_DEPLOYMENT=gpt-4o go run ./examples/coding_agent/

# Google Cloud Vertex AI
GOOGLE_CLOUD_PROJECT=my-project GOOGLE_ACCESS_TOKEN=$(gcloud auth print-access-token) \
  go run ./examples/coding_agent/

Demonstrates:

  • Wiring built-in file tools + shell tools + custom tools onto one agent
  • VectorKnowledge with in-memory embeddings for code search (RAG)
  • An autonomous agent loop bounded by MaxIterations
  • Combining tools with system prompts for effective coding workflows
  • Any provider (OpenAI, Azure OpenAI, Vertex AI, Bedrock, …); mock fallback when nothing is set

multi_agent​

All 4 team strategies (sequential, parallel, router, coordinator), direct channels, and bus delegation. Works with a mock provider if no API key is set.

# With mock (no key)
go run ./examples/multi_agent/

# With OpenAI
OPENAI_API_KEY=sk-... go run ./examples/multi_agent/

# With Azure OpenAI
AZURE_OPENAI_API_KEY=... AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com \
  AZURE_OPENAI_DEPLOYMENT=gpt-4o go run ./examples/multi_agent/

# With Google Cloud Vertex AI
GOOGLE_CLOUD_PROJECT=my-project GOOGLE_ACCESS_TOKEN=$(gcloud auth print-access-token) \
  go run ./examples/multi_agent/

Demonstrates:

  • Sequential, parallel, router, and coordinator team strategies
  • Direct agent-to-agent channels and inter-agent bus delegation
  • Any provider (OpenAI, Azure OpenAI, Vertex AI, Bedrock, …); mock fallback when nothing is set

team_deploy​

Deploy multi-agent teams from YAML with sandbox isolation. Loads a team deployment config, builds agents with YAML-defined tools, and runs them in a sandboxed process environment using sequential and coordinator strategies.

# OpenAI
OPENAI_API_KEY=sk-... go run ./examples/team_deploy/

# Azure OpenAI
AZURE_OPENAI_API_KEY=... AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com \
  AZURE_OPENAI_DEPLOYMENT=gpt-4o go run ./examples/team_deploy/

# Google Cloud Vertex AI
GOOGLE_CLOUD_PROJECT=my-project GOOGLE_ACCESS_TOKEN=$(gcloud auth print-access-token) \
  go run ./examples/team_deploy/

# Or via the CLI:
chronos deploy examples/team_deploy/deploy.yaml "Add error handling to the API"

Demonstrates:

  • YAML-driven agent and team configuration
  • Sandbox-backed tool execution for safe agent autonomy
  • Sequential pipeline vs. coordinator strategies
  • Deploying a full coding team from a single config file
  • Any provider (OpenAI, Azure OpenAI, Vertex AI, Bedrock, …); mock fallback when nothing is set