Add AG2 (formerly AutoGen) multi-agent examples

ag2-agents/payment-approval/.env.example

@@ -0,0 +1,5 @@
+OPENAI_API_KEY=your-openai-api-key-here
+
+# Optional
+LLM_MODEL=gpt-4o-mini
+OPENAI_BASE_URL=https://api.openai.com/v1

ag2-agents/payment-approval/README.md

@@ -0,0 +1,26 @@
+# AG2 Two-Agent Payment Approval
+
+Demonstrates AG2's `human_input_mode="ALWAYS"` pattern as an approval gate
+before triggering a Skyfire payment — the first example in this repo requiring
+explicit user confirmation before a financial action.
+
+Two agents collaborate:
+- **researcher** — investigates the recipient and produces a risk assessment
+- **payment_executor** — presents the assessment, pauses for human confirmation,
+  then executes or aborts based on the response
+
+## Key AG2 Features
+
+- **`human_input_mode="ALWAYS"`** — executor pauses before every response; human
+  types "yes" to proceed or "no" to abort — no custom routing logic needed
+- **Two-agent `initiate_chat`** — researcher hands off to executor via the
+  natural conversation flow; the shared message history carries the assessment
+
+## Quick Start
+
+```bash
+cd ag2-agents/payment-approval
+pip install -r requirements.txt
+cp .env.example .env
+python main.py
+```

ag2-agents/payment-approval/main.py

@@ -0,0 +1,70 @@
+"""
+AG2 two-agent payment approval with human-in-the-loop gate.
+
+researcher — investigates the recipient and produces a risk assessment
+executor   — presents the assessment and waits for explicit human confirmation
+             before executing the Skyfire payment
+
+human_input_mode="ALWAYS" on executor is the approval gate: the agent
+pauses before every response so the human can type "yes" to proceed or
+"no" to abort. No custom routing logic required.
+"""
+import os
+from dotenv import load_dotenv
+from autogen import ConversableAgent, LLMConfig
+
+load_dotenv()
+
+llm_config = LLMConfig(
+    {
+        "model": os.getenv("LLM_MODEL", "gpt-4o-mini"),
+        "api_key": os.environ["OPENAI_API_KEY"],
+        "base_url": os.getenv("OPENAI_BASE_URL", "https://api.openai.com/v1"),
+    },
+    temperature=0.2,
+    cache_seed=None,
+)
+
+researcher = ConversableAgent(
+    name="researcher",
+    system_message=(
+        "You are a payment risk analyst. Investigate the payment recipient using available "
+        "tools: check their Fetch.ai address history, reputation, and any known flags. "
+        "Produce a concise risk assessment with a clear recommendation (proceed / do not proceed). "
+        "End your assessment with ASSESSMENT COMPLETE."
+    ),
+    llm_config=llm_config,
+    is_termination_msg=lambda m: "ASSESSMENT COMPLETE" in (m.get("content") or ""),
+)
+
+executor = ConversableAgent(
+    name="payment_executor",
+    system_message=(
+        "You handle payment execution. Present the researcher's risk assessment clearly, "
+        "state the exact payment details (recipient, amount, reason), then ask the human "
+        "to confirm. If the human approves, call the skyfire_send tool to execute the payment. "
+        "If the human declines, acknowledge and terminate. End with TERMINATE."
+    ),
+    llm_config=llm_config,
+    human_input_mode="ALWAYS",  # pauses before every response — the human types yes/no
+    is_termination_msg=lambda m: "TERMINATE" in (m.get("content") or ""),
+)
+
+
+def run_payment_approval(recipient: str, amount: float, reason: str) -> None:
+    researcher.initiate_chat(
+        executor,
+        message=(
+            f"Payment request: {amount} USDC to {recipient} — reason: '{reason}'. "
+            f"Investigate the recipient and produce a risk assessment."
+        ),
+        max_turns=6,
+    )
+
+
+if __name__ == "__main__":
+    run_payment_approval(
+        recipient="alice.fetch",
+        amount=50.0,
+        reason="research report delivery",
+    )

ag2-agents/payment-approval/requirements.txt

@@ -0,0 +1,2 @@
+ag2[openai]>=0.11.0
+python-dotenv>=1.0.0

ag2-agents/payment-approval/tests/conftest.py

@@ -0,0 +1,9 @@
+"""
+Ensure payment-approval/ is on sys.path so 'import main' works in tests.
+"""
+import sys
+import os
+
+parent = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+if parent not in sys.path:
+    sys.path.insert(0, parent)

ag2-agents/payment-approval/tests/test_payment.py

@@ -0,0 +1,23 @@
+import os
+os.environ.setdefault("OPENAI_API_KEY", "test-key")
+
+def test_agents_instantiate():
+    """Verify agent setup without initiating a chat."""
+    import main as m
+    assert m.researcher.name == "researcher"
+    assert m.executor.name == "payment_executor"
+
+def test_executor_human_input_mode():
+    """Executor must have human_input_mode=ALWAYS — the approval gate."""
+    import main as m
+    assert m.executor.human_input_mode == "ALWAYS"
+
+def test_researcher_termination_condition():
+    import main as m
+    assert m.researcher._is_termination_msg({"content": "Risk: low. ASSESSMENT COMPLETE"}) is True
+    assert m.researcher._is_termination_msg({"content": "Still investigating..."}) is False
+
+def test_executor_termination_condition():
+    import main as m
+    assert m.executor._is_termination_msg({"content": "Payment aborted. TERMINATE"}) is True
+    assert m.executor._is_termination_msg({"content": "Please confirm."}) is False

ag2-agents/research-synthesis-team/.env.example

@@ -0,0 +1,12 @@
+OPENAI_API_KEY=your-openai-api-key-here
+AGENTVERSE_API_KEY=your-agentverse-api-key-here
+
+# Optional
+LLM_MODEL=gpt-4o-mini
+OPENAI_BASE_URL=https://api.openai.com/v1
+AGENT_PORT=8008
+AGENTVERSE_URL=https://agentverse.ai
+
+# Leave empty to use DuckDuckGo search (default, no API key needed).
+# Set to a Fetch.ai MCP gateway URL to use MCP tools instead.
+MCP_SERVER_URL=

ag2-agents/research-synthesis-team/Dockerfile

@@ -0,0 +1,14 @@
+FROM python:3.11-slim
+
+WORKDIR /app
+
+COPY . .
+
+RUN apt-get update && apt-get install -y gcc \
+    && pip install --no-cache-dir --upgrade pip \
+    && pip install --no-cache-dir -r requirements.txt \
+    && apt-get clean && rm -rf /var/lib/apt/lists/*
+
+ENV PYTHONUNBUFFERED=1
+EXPOSE 8008
+CMD ["python", "main.py"]

ag2-agents/research-synthesis-team/README.md

@@ -0,0 +1,44 @@
+# AG2 Research Synthesis Team
+
+![ag2](https://img.shields.io/badge/ag2-00ADD8) ![uagents](https://img.shields.io/badge/uagents-4A90E2) ![a2a](https://img.shields.io/badge/a2a-000000) ![innovationlab](https://img.shields.io/badge/innovationlab-3D8BD3)
+
+A multi-agent research pipeline using [AG2](https://github.com/ag2ai/ag2) (formerly AutoGen)
+integrated with the Fetch.ai uAgents ecosystem via the A2A protocol.
+
+## Architecture
+
+Four specialists collaborate under GroupChat with LLM-driven speaker selection, wrapped as
+an A2A executor and exposed as a discoverable agent on Agentverse.
+
+```
+User / ASI:One / other uAgent
+         ↓
+SingleA2AAdapter (port 8008) → Agentverse
+         ↓
+AG2ResearchExecutor (A2A AgentExecutor)
+         ↓
+GroupChat (AG2)
+├── web_researcher   — DuckDuckGo search, gathers sources
+├── financial_analyst — market data, metrics, trends
+├── tech_analyst     — technical feasibility, risks
+└── synthesizer      — final structured report
+```
+
+## Quick Start
+
+```bash
+pip install -r requirements.txt
+cp .env.example .env  # add OPENAI_API_KEY and AGENTVERSE_API_KEY
+python main.py
+```
+
+No additional API keys needed for search — DuckDuckGo is used by default.
+
+To use a Fetch.ai MCP gateway instead, set `MCP_SERVER_URL` in `.env`.
+
+## AG2 Features Demonstrated
+
+- **`GroupChat` with `speaker_selection_method="auto"`** — LLM-driven dynamic speaker selection
+- **`DuckDuckGoSearchTool`** — built-in web search, no API key required
+- **Native MCP client** — optional override via `MCP_SERVER_URL` for richer tool access
+- **`A2A AgentExecutor`** — same integration pattern used by other examples in this repo

ag2-agents/research-synthesis-team/agent_executor.py

@@ -0,0 +1,47 @@
+"""
+Wraps the AG2 GroupChat workflow as an A2A AgentExecutor
+for use with SingleA2AAdapter (Pattern B).
+"""
+from a2a.server.agent_execution import AgentExecutor, RequestContext
+from a2a.server.events import EventQueue
+from a2a.types import Part, TextPart
+from a2a.utils import new_agent_text_message
+from autogen import LLMConfig
+from typing_extensions import override
+
+from workflow import run_research
+
+
+class AG2ResearchExecutor(AgentExecutor):
+    """A2A-compatible executor wrapping the AG2 GroupChat workflow."""
+
+    def __init__(self, llm_config: LLMConfig, mcp_url: str | None = None):
+        self.llm_config = llm_config
+        self.mcp_url = mcp_url
+
+    @override
+    async def execute(self, context: RequestContext, event_queue: EventQueue) -> None:
+        message_content = ""
+        for part in context.message.parts:
+            if isinstance(part, Part):
+                if isinstance(part.root, TextPart):
+                    message_content = part.root.text
+                    break
+
+        if not message_content:
+            await event_queue.enqueue_event(
+                new_agent_text_message("Error: No message content received.")
+            )
+            return
+
+        try:
+            result = await run_research(message_content, self.llm_config, self.mcp_url)
+            await event_queue.enqueue_event(new_agent_text_message(result))
+        except Exception as e:
+            await event_queue.enqueue_event(
+                new_agent_text_message(f"Research failed: {e}")
+            )
+
+    @override
+    async def cancel(self, context: RequestContext, event_queue: EventQueue) -> None:
+        raise Exception("Cancel not supported for this agent executor.")

ag2-agents/research-synthesis-team/agents.py

@@ -0,0 +1,84 @@
+"""
+AG2 (formerly AutoGen) research synthesis team.
+Four specialists collaborate under GroupChat with LLM-driven speaker selection.
+"""
+from autogen import AssistantAgent, LLMConfig
+
+
+def build_agents(llm_config: LLMConfig) -> list[AssistantAgent]:
+    web_researcher = AssistantAgent(
+        name="web_researcher",
+        system_message=(
+            "You are a web research specialist with access to search tools.\n\n"
+            "WORKFLOW:\n"
+            "1. Use the search tool to find relevant sources on the assigned topic.\n"
+            "2. Run at least 2-3 different search queries to cover the topic broadly.\n"
+            "3. Compile your findings into a structured response.\n\n"
+            "OUTPUT FORMAT (mandatory):\n"
+            "- Provide a markdown table of sources found:\n"
+            "  | # | Title | URL | Key Finding |\n"
+            "- Minimum 5 sources with direct links.\n"
+            "- Below the table, write a 200-word summary of the most important findings.\n"
+            "- Flag any conflicting information between sources."
+        ),
+        llm_config=llm_config,
+    )
+    financial_analyst = AssistantAgent(
+        name="financial_analyst",
+        system_message=(
+            "You are a financial analyst. Analyse market data, trends, and economic "
+            "implications of the research topic using the sources provided by the "
+            "web researcher.\n\n"
+            "OUTPUT FORMAT (mandatory):\n"
+            "- Provide a financial summary table:\n"
+            "  | Metric | Value | Source | Trend |\n"
+            "- Include: market size, growth rate, key players, funding, and revenue "
+            "  figures where available. Use 'N/A' when data is unavailable.\n"
+            "- Write a 150-word analysis of financial risks and opportunities.\n"
+            "- Be quantitative — use numbers, percentages, and dollar amounts."
+        ),
+        llm_config=llm_config,
+    )
+    tech_analyst = AssistantAgent(
+        name="tech_analyst",
+        system_message=(
+            "You are a technology analyst. Evaluate the technical landscape of the "
+            "research topic using the sources provided by the web researcher.\n\n"
+            "OUTPUT FORMAT (mandatory):\n"
+            "- Provide a technology assessment table:\n"
+            "  | Technology/Component | Maturity | Adoption | Risk Level | Notes |\n"
+            "- Maturity values: Emerging, Growing, Mature, Declining.\n"
+            "- Risk Level: Low, Medium, High.\n"
+            "- Write a 150-word analysis covering: technical feasibility, innovation "
+            "  potential, and key technical challenges.\n"
+            "- Identify the top 3 technical risks with mitigation strategies."
+        ),
+        llm_config=llm_config,
+    )
+    synthesizer = AssistantAgent(
+        name="synthesizer",
+        system_message=(
+            "You are a synthesis expert. Once all specialists have contributed, "
+            "produce a final structured report combining all perspectives.\n\n"
+            "MANDATORY REPORT STRUCTURE:\n"
+            "## Executive Summary\n"
+            "3-5 bullet points covering the most important findings.\n\n"
+            "## Research Findings\n"
+            "Key sources and discoveries from the web researcher.\n\n"
+            "## Financial Analysis\n"
+            "Market data, financial metrics, and economic outlook.\n\n"
+            "## Technical Analysis\n"
+            "Technology landscape, maturity assessment, and risks.\n\n"
+            "## Conclusions & Recommendations\n"
+            "3-5 actionable recommendations ranked by priority.\n\n"
+            "## Sources\n"
+            "Consolidated list of all sources cited, as [Title](URL).\n\n"
+            "RULES:\n"
+            "- Do NOT repeat raw data from specialists — synthesize and add insight.\n"
+            "- Total report length: 500-800 words.\n"
+            "- End your response with TERMINATE."
+        ),
+        llm_config=llm_config,
+        is_termination_msg=lambda m: "TERMINATE" in (m.get("content") or ""),
+    )
+    return [web_researcher, financial_analyst, tech_analyst, synthesizer]

ag2-agents/research-synthesis-team/main.py

@@ -0,0 +1,43 @@
+"""
+Fetch.ai uAgent exposing the AG2 research team via A2A protocol (Pattern B).
+Discoverable on Agentverse; callable from ASI:One or other uAgents.
+"""
+import os
+from dotenv import load_dotenv
+from uagents_adapter import SingleA2AAdapter
+from autogen import LLMConfig
+
+from agent_executor import AG2ResearchExecutor
+
+load_dotenv()
+
+llm_config = LLMConfig(
+    {
+        "model": os.getenv("LLM_MODEL", "gpt-4o-mini"),
+        "api_key": os.getenv("OPENAI_API_KEY", ""),
+        "base_url": os.getenv("OPENAI_BASE_URL", "https://api.openai.com/v1"),
+    },
+    temperature=0.3,
+    cache_seed=None,
+)
+
+executor = AG2ResearchExecutor(
+    llm_config=llm_config,
+    mcp_url=os.getenv("MCP_SERVER_URL"),  # optional: Fetch.ai MCP gateway
+)
+
+adapter = SingleA2AAdapter(
+    agent_executor=executor,
+    name="AG2 Research Synthesis Team",
+    description=(
+        "Multi-agent research team using AG2 (formerly AutoGen). "
+        "Four specialists (web researcher, financial analyst, tech analyst, synthesizer) "
+        "collaborate to produce comprehensive research reports on any topic."
+    ),
+    port=int(os.getenv("AGENT_PORT", "8008")),
+    agentverse_url=os.getenv("AGENTVERSE_URL", "https://agentverse.ai"),
+    mailbox_api_key=os.getenv("AGENTVERSE_API_KEY", ""),
+)
+
+if __name__ == "__main__":
+    adapter.run()

ag2-agents/research-synthesis-team/requirements.txt

@@ -0,0 +1,6 @@
+ag2[openai,mcp,duckduckgo]>=0.11.0
+a2a-sdk
+mcp>=1.0.0
+uagents>=0.20.0
+uagents-adapter>=0.4.0
+python-dotenv>=1.0.0