Agent Stack Showcase Agent 🤖💬🧪

October 30, 2025 · View on GitHub

The Agent Stack Showcase Agent is a research prototype built with the BeeAI Framework and Agent Stack SDK.

It demonstrates how to combine tool orchestration, memory, file analysis, and platform extensions into a general-purpose conversational assistant. The agent can handle chat, process uploaded files, search the web, and provide structured outputs with citations and trajectory logs for debugging and UI replay.

✨ Capabilities

  • Multi-turn chat with persistent per-session memory (UnconstrainedMemory)

  • Tool orchestration via the experimental RequirementAgent, with rules like:

    • ThinkTool — invoked first and after every tool for reasoning
    • DuckDuckGoSearchTool — used up to 2 times per query, skipped for casual greetings
    • File processing — supports PDF, CSV, JSON, and plain text uploads

  • Citation extraction — converts [text](url) markdown links into structured citation objects

  • Trajectory tracking — logs each reasoning step, tool invocation, and output for replay/debugging

  • Configurable settings — users can toggle thinking/search behaviors and select response style (concise, standard, detailed)

  • Basic error handling — user-facing messages and detailed logs

🚀 Running the Agent

  1. Install Agent Stack Follow the Quickstart Guide to install and set up Agent Stack. This is required before running the agent.

  2. Start the server Once the platform is installed, launch the agent server:

    uv run server

    The server runs on the configured HOST and PORT environment variables (defaults: 127.0.0.1:8000).

🧩 Key Components

  • agentstack_showcase(...) — Main async entrypoint handling chat, file uploads, memory, and tool orchestration
  • RequirementAgent(...) — Experimental agent that enforces ConditionalRequirement rules for tool usage
  • ThinkTool — Provides structured reasoning and analysis
  • DuckDuckGoSearchTool — Performs real-time web search (with constraints)
  • extract_citations(...) — Converts markdown links into structured citation objects
  • is_casual(...) — Skips tool invocation for short greetings or casual input
  • get_memory(...) — Provides per-session UnconstrainedMemory
  • run() — Starts the Agent Stack server

🔌 Extensions

  • CitationExtensionServer — renders citations into structured previews
  • TrajectoryExtensionServer — captures reasoning/tool usage for UI replay & debugging
  • LLMServiceExtensionServer — manages LLM fulfillment through Agent Stack
  • SettingsExtensionServer — allows user configuration of agent behaviors

💡 Example Interaction

User input:

What are the latest advancements in AI research from 2025?

Agent flow:

  1. ThinkTool invoked for reasoning
  2. DuckDuckGoSearchTool called (unless skipped for casual input)
  3. Response returned with proper [label](url) citations
  4. Citations extracted and sent to UI
  5. Steps logged in trajectory extension
  6. Conversation context persisted for future turns
  7. If a file is uploaded, it’s analyzed and summarized

📂 Example Skills

The agent supports both chat and file analysis, such as:

  • "What are the latest advancements in AI research from 2025?"
  • "Can you help me write a Slack announcement for [topic/team update]?"
  • "Analyze this CSV file and tell me the key trends."
  • "Summarize the main points from this PDF document."