Build an AI Agent That Runs 24/7 With Tank

Build an AI Agent That Runs 24/7 With Tank

Building AI Agents for 24/7 Operations

Introduction to AI Agents

  • The speaker discusses the vision of having AI agents that operate continuously, enhancing productivity and quality of life.
  • Examples include content creation, health monitoring through wearables, and tracking price changes on e-commerce platforms.

Updates on Tank

  • A new version of Tank has been deployed, featuring automatic updates and scheduled database backups.
  • The importance of database backups is highlighted to prevent operational disruptions in the event of an update failure.
  • Introduction of PI agent support and full file system access for Codex is mentioned as a significant advancement.

Understanding Tank

  • Tank is described as an AI coding orchestration system, similar to Claude code and Codex.
  • It allows users to perform web searches, utilize tools, and generate detailed reports efficiently.

Personal Experience with Tank

  • The speaker shares their personal projects within Tank, including an AI video editor named Kino.
  • Various agents can be integrated into Tank for enhanced functionality.

Transitioning to Using Tank

Evolution from Other Tools

  • The speaker recounts their journey from using OpenClaw and Hermes to fully adopting Tank for all operations.
  • Initially built in response to limitations imposed by Anthropic on Mac usage for CLAUDE code.

Setting Up a Fresh Instance

  • Demonstration of setting up a fresh instance of Tank without pre-existing agents or configurations.

Configuring Agents in Tank

Agent Integration Process

  • Users can easily integrate various coding agents like Claude code into their projects within the settings menu.

Authentication Steps

  • Signing into Claude code requires authentication via OAuth; this process is demonstrated live.

Creating Projects in Tank

Project Initialization

  • A new project titled "Second Brain" is created as part of the demonstration process.

Features Overview

  • Discussion about local AI features available when connected properly; emphasizes user control over data management.

Building Scheduled Tasks

Task Creation Process

  • The speaker initiates creating a task that instructively summarizes top posts from Reddit's Singularity subreddit.

Challenges Faced

  • Initial attempts at accessing Reddit face obstacles due to bot restrictions; alternative methods are explored.

Learning from Execution

Task Performance Analysis

  • After several attempts, the task successfully retrieves information but highlights inefficiencies due to website restrictions.

Knowledge Retention Strategies

  • Emphasizes saving learned strategies either as markdown files or skills within the project for future efficiency improvements.

Enhancing Efficiency with Skills

Skill Development

  • By creating specific skills based on previous tasks (e.g., fetching Reddit posts), subsequent executions become significantly faster.

Conclusion on Scheduled Tasks

  • Final thoughts focus on building an agent capable of running tasks autonomously around the clock while improving its performance over time through learned experiences.

What is a Cron Job and How to Schedule Tasks?

Understanding Cron Jobs

  • A cron job is essentially a scheduled task that runs at predetermined times, allowing for automation of processes.
  • Openclaw refers to this scheduling feature as "cron," similar to how Hermes Agent manages scheduled tasks.

Setting Up Scheduled Tasks

  • Users can set up tasks with specific timings, such as daily or hourly, tailored to their local time zones.
  • The current maximum frequency for running jobs is every hour, but future updates may allow for more frequent executions.

Task Instructions and Notifications

  • Users can specify detailed instructions for the tasks, such as searching subreddits and summarizing posts.
  • Notification settings are customizable; currently supporting Google Chat and NTFY, with potential expansions in the future.

Creating Multiple Scheduled Tasks

Scheduling Additional Tasks

  • Users can create multiple tasks simultaneously; for example, fetching summaries from various AI-related subreddits.
  • The system allows users to synthesize information from different sources into a cohesive summary based on recent trends.

Utilizing Different Models

  • There’s no need to rely solely on expensive models like Fable or Sonnet; smaller models like Haiku 4.5 can be effective for certain tasks.

Real-Time Execution of Scheduled Tasks

Monitoring Task Execution

  • Once scheduled, users can observe the execution of their tasks in real-time through the interface.
  • The system enables hands-off operation where agents autonomously gather information overnight.

Results and Insights from Executions

  • Upon completion of tasks, results are displayed showing dominant topics discussed across selected subreddits.

Evaluating Model Performance

Comparing Outputs from Different Models

  • Users can verify outputs by cross-referencing results against subreddit data to ensure accuracy and relevance.

Addressing Hallucinations in AI Responses

  • Discussions highlight concerns about AI-generated content potentially hallucinating facts or misrepresenting data.

Integrating New Agents and Features

Adding New Agents

  • Efforts are underway to integrate additional agents like Grok into the existing framework for enhanced functionality.

Community Engagement

  • Viewers are encouraged to join the community around Tank for collaborative development and support in building AI agents.

Troubleshooting Installation Issues

Installation Challenges

  • During live demonstrations, issues arise with installing new agents due to fresh setups which require configuration adjustments.

User Interaction During Setup

  • Viewers are invited to engage during setup processes while troubleshooting installation challenges collaboratively.

Final Thoughts on Automation Potential

Expanding Use Cases

  • The potential applications of automated agents include health monitoring reports, news synthesis, email management, etc., showcasing vast possibilities within personal productivity enhancements.

Exploring AI Agents and Scheduled Tasks

Introduction to Anthropic News

  • The discussion begins with a mention of synthesized anthropic news, highlighting the integration of Reddit posts for accurate information retrieval.
  • Emphasis on the legality and accuracy of using upvoted content from Reddit, showcasing Grok's ability to fetch reliable data.

Functionality of AI Agents

  • A scenario is presented where a single trusted agent could relay information without sharing personal email credentials, enhancing privacy.
  • The potential for one agent to call another upon task completion is introduced, utilizing the Tank API for seamless communication between agents.

Task Scheduling and Optimization

  • Demonstration of editing tasks in Anthropic Daily to send updates via Tank API, illustrating practical applications of scheduled tasks.
  • Discussion on optimizing task durations and success rates, with aspirations for visual representations like graphs to track performance over time.

Real-Time Data Retrieval

  • Highlighting current top posts on artificial intelligence from Reddit, including user reactions that provide context beyond just headlines.
  • The importance of understanding public sentiment alongside news stories is emphasized as a valuable feature.

Future Improvements in Agent Functionality

  • Current limitations are acknowledged regarding how often agents can run tasks (once every hour), but excitement about future enhancements is expressed.
  • Anticipation builds around the potential improvements in agent capabilities that could lead to more efficient workflows.

Community Engagement and User Experiences

Live Interaction with Viewers

  • Acknowledgment of viewers tuning into the live stream, fostering community engagement through real-time interaction.
  • Sharing user experiences where AI identifies songs from radio stations demonstrates practical applications of technology in everyday life.

Experimenting with Grok's Capabilities

  • Curiosity about Grok’s ability to access data from X (formerly Twitter), indicating ongoing exploration of its functionalities.
  • Initiating a new task within Grok to summarize recent posts from a specific account showcases hands-on experimentation with AI tools.

Assessing Accuracy and Performance

Testing Data Retrieval Features

  • Running tests on Grok’s capability to retrieve latest posts without relying on paid services highlights its versatility.
  • Observations about Fable 5's performance raise questions about its effectiveness compared to other solutions available in the market.

Video Transcript Extraction

  • Inquiry into whether Grok can extract transcripts from videos posted on X indicates an interest in expanding its functionality further.
  • Successful extraction of video transcripts would significantly enhance Grok's utility, marking an important milestone if achieved.

Conclusion and Future Directions

Final Thoughts Before Community Engagement

  • As preparations are made for joining the community live session, reflections on various discussions highlight ongoing developments within AI systems.
  • Insights into technical challenges faced during streaming emphasize the trade-offs between quality and real-time delivery.

Closing Remarks

  • The successful retrieval of video content marks a significant achievement for Grok; anticipation builds around future capabilities as community interactions continue.
Video description

Get Tank - join the community 👉 https://mrc.fm/cmc Most people run an AI prompt, get an answer and close the tab. But the real magic is an agent that keeps working when you don't: pulling the news into a daily blog draft, watching listings, running scheduled jobs all night while you sleep. In this stream I'll build one live with Tank, a self-hosted AI worker that lives on your own box and never leaves your data behind. We'll wire it to a schedule, hand it a real recurring task, and cover the bit everyone asks about: how to set it, trust it and walk away without babysitting a terminal. If you've ever wanted your own AI employee that just gets on with it, this is the build. 0:00 Building autonomous AI agents 1:01 Tank orchestration system updates 2:11 How Tank orchestrates coding agents 4:03 Replacing OpenClaw and Hermes 5:20 Connecting Claude Code with OAuth 7:32 Creating a Second Brain project 11:53 Scraping Reddit with Claude Code 18:52 Saving agent skills in Tank 24:07 Scheduling cron jobs for agents 29:51 Running scheduled tasks in real-time 34:08 Connecting Grok Build to Tank 42:39 Split testing Grok and Claude 47:19 Scheduling advanced sentiment analysis 54:10 Scraping X posts with Grok 56:47 Extracting video transcripts using Grok