Self-learning poker playing agent

[rudy-pants]

Project link

https://github.com/rudy-pants/temporal-agent-poker/tree/main

Language

Python

Short description (max 256 chars)

Suppose you wanted a demonstration of a self-learning agent that could improve it's performance by updating it's own memory. This is what our project showcases, in a direct poker game between a human player, and an agent capable of improving itself to superhuman levels of play.

Long Description

Our project contains 3 major parts: a frontend client, a temporal worker running all backend workflows, and a memory module for agent self-learning. The agent will collect observations about the game each hand, and then compress those into memories after each game, updating it's memories and playstyle at it's own discretion.

https://drive.google.com/file/d/1kfkVUa9E40_iHv0vj2QlXWRP7ZOg2i22/view?usp=sharing

Author(s)

Rudy Banerjee, Geico
Khalid Shaikh, Apple

[github-actions]

Hi, I'm ZiggyBot! 🤖 Here's my pre-evaluation of this submission:

---

Criterion	Assessment
Useful to Temporal users	✅ Yes — This project demonstrates a compelling real-world pattern: orchestrating an AI agent with persistent memory and self-learning capabilities using Temporal workflows, which is highly relevant for developers building agentic applications.
Benefits clearly explained	✅ Yes — The README includes an excellent architecture diagram and clearly explains how Temporal features (signals, queries, timers, activities, continue-as-new, replay) are leveraged for game orchestration and agent memory management.
OSI Approved License	❌ No — GitHub reports "Not detected" for the license, meaning there's no LICENSE file in the repository.
README with working instructions	⚠️ Unclear — The README shows a detailed architecture diagram but the provided content was truncated, so I cannot verify whether complete setup/installation instructions are present.

Overall: Needs review ⚠️

Notes:

• This is a genuinely creative and well-architected project that showcases multiple advanced Temporal patterns working together: signals for player input, queries for game state observation, timers for turn timeouts, activities for LLM calls, and continue-as-new for multi-hand sessions with bounded history.
• The self-learning memory system where the agent reflects after each hand and compresses learnings to persistent storage is a thoughtful demonstration of durable agent memory—a pattern with broad applicability beyond poker.
• The missing license is a blocker for Code Exchange acceptance, but it's an easy fix!
• The video demo link is a nice touch for understanding the project in action.

Suggested questions for submitter:

• Could you add an OSI-approved license (MIT recommended) to the repository?
• Is there a complete "Getting Started" or installation section in your README? The truncated view didn't show setup steps—confirming these exist would help.
• Have you considered adding any tests or example scenarios that demonstrate the agent's learning progression over multiple games?

Teaching moment:

This project beautifully demonstrates using continue_as_new to manage long-running game sessions with bounded event history while preserving agent state across hands. The workflow processes each hand, runs a reflection activity, then continues as new—allowing multi-hand poker sessions without unbounded history growth. The combination of durable timers for turn timeouts (surviving worker crashes), signals for human player actions, and activities for LLM decision-making shows how Temporal can orchestrate complex human-AI interaction loops reliably.
View in source ↗

---

---

ZiggyBot is an AI pre-screener based on Temporal's community mascot Ziggy. Final decisions are made by the community team.

Reviewer: Check the box below to generate a Contentful-ready long description for this submission.

• Generate long description if needed

[github-actions]

Hi, I'm ZiggyBot! 🤖 Here's my pre-evaluation of this submission:

---

Criterion	Assessment
Useful to Temporal users	✅ Yes — Demonstrates a compelling AI agent pattern with durable state management, self-learning memory, and real-time human interaction via signals/queries.
Benefits clearly explained	✅ Yes — The README's architecture diagram and detailed flow explanation clearly show how Temporal primitives (signals, queries, timers, continue-as-new) enable durable game orchestration.
OSI Approved License	❌ No — GitHub reports "Not detected" and no LICENSE file is visible in the repository.
README with working instructions	⚠️ Unclear — The README shows architecture but appears truncated; setup instructions (dependencies, environment variables, how to run) may be incomplete or missing.

Overall: Needs review ⚠️

Notes:

• This is a genuinely creative showcase of Temporal for AI agent orchestration! The architecture is thoughtful, demonstrating signals for player actions, queries for game state, durable timers for turn timeouts, and continue-as-new for multi-hand sessions.
• The self-learning memory system (filesystem-based persistence with GPT reflection) is an interesting pattern that could inspire other agent-based Temporal projects.
• The claim of "superhuman levels of play" is ambitious marketing—worth softening to something more accurate like "adaptive" or "improving."
• Missing license is a blocker for Code Exchange inclusion; this is an easy fix!
• Would love to see the full README with setup instructions to verify completeness.

Suggested questions for submitter:

• Could you add an OSI-approved license (MIT recommended) to the repository?
• Is there a complete "Getting Started" section with setup instructions (installing dependencies, configuring OpenAI API key, starting Temporal, running the worker and frontend)?
• The video link is a Google Drive file—would you consider uploading a demo to YouTube or embedding a GIF in the README for easier access?

Teaching moment:

This project demonstrates an excellent use of continue_as_new for long-running game sessions with bounded history. After each hand, the workflow reflects on results and then continues as new, carrying forward compressed state (running notes, chip counts) while keeping the event history manageable. This pattern is essential for any Temporal workflow that could run indefinitely—games, monitoring agents, or chat sessions. The per-hand reflection activity followed by continue-as-new shows how to balance durable state with practical history limits.
View in source ↗

---

---

ZiggyBot is an AI pre-screener based on Temporal's community mascot Ziggy. Final decisions are made by the community team.

Reviewer: Check the box below to generate a Contentful-ready long description for this submission.

• Generate long description if needed

[github-actions]

Hi, I'm ZiggyBot! 🤖 Here's my pre-evaluation of this submission:

---

Criterion	Assessment
Useful to Temporal users	✅ Yes — Demonstrates a sophisticated agentic AI pattern with self-learning memory, signals for human input, queries for state observation, and continue-as-new for bounded history—all orchestrated elegantly through Temporal workflows.
Benefits clearly explained	✅ Yes — The README includes an excellent architecture diagram and clearly explains how Temporal's features (signals, queries, timers, activities, replay, continue-as-new) enable durable game state and AI decision-making.
OSI Approved License	❌ No — No license file was detected by GitHub, which means the project cannot be redistributed or used by others under clear legal terms.
README with working instructions	⚠️ Unclear — The README was truncated but the architecture section is thorough; however, it's not clear if there are complete setup/installation instructions including environment variables (e.g., OpenAI API key) and dependency installation steps.

Overall: Needs review ⚠️

Notes:

• This is a genuinely creative and well-architected project! The self-learning agent pattern with filesystem-based memory persistence is a compelling demonstration of agentic workflows.
• The use of continue-as-new after each hand to manage history size while maintaining running notes across hands shows thoughtful design.
• The combination of human interaction via signals with 30-second timeouts and GPT-powered AI decisions via activities is a great real-world pattern.
• The project would benefit greatly from adding a license (MIT recommended) and ensuring the README has complete "Getting Started" instructions.
• Zero stars suggests this is a new submission—exciting to see fresh community contributions!

Suggested questions for submitter:

• Would you be willing to add an MIT (or other OSI-approved) license to the repository?
• Could you confirm the README includes complete setup instructions, particularly around installing dependencies, setting up the OpenAI API key, and starting all required services (Temporal, FastAPI, React frontend)?
• The video link is a Google Drive share—would you consider hosting it on YouTube or embedding a GIF in the README for easier access?

Teaching moment:

This project showcases a beautiful pattern for human-in-the-loop workflows with AI fallback. The workflow uses wait_condition with a 30-second timer for human input via signals, and if the human doesn't respond, it falls back to an AI activity. This is a reusable pattern for any interactive application where you want durable timeouts that survive crashes. The memory system that persists learnings to the filesystem and loads them into the GPT system prompt at session start demonstrates how Temporal activities can bridge stateless AI calls with persistent state.
View in source ↗

---

---

ZiggyBot is an AI pre-screener based on Temporal's community mascot Ziggy. Final decisions are made by the community team.

Reviewer: Check the box below to generate a Contentful-ready long description for this submission.

• Generate long description if needed