1011: tmux + Terminal Maxxing with Ben Vinegar

The podcast discusses the integration of AI agents with terminal environments, emphasizing tools like Tmux for managing multiple tasks and sessions, though skepticism about its hype is noted. Tools such as Caffeinate (to keep devices awake) and VPS providers like Hetzner or DigitalOcean are highlighted for running agents continuously. The discussion also covers challenges with agent autonomy, including risks of uncontrolled task execution (e.g., generating GitHub issues without oversight) and the need for systems like JIRA to prioritize tasks and enforce tool approval blocks. Infrastructure considerations include portable setups with battery-powered devices and internet access, as well as risks like agent stalling due to unapproved actions or hardware sleep states.

A significant focus is on the Modem project, an AI-driven tool for automating non-coding tasks like synthesizing feedback and managing ticket backlogs, with a thesis that AI will dominate coding workflows, leaving non-coding tasks as the new bottleneck. Terminal-based workflows are explored in depth, including the use of Tmux for splitting panes, detaching sessions, and integrating with coding agents (e.g., Kimmy, Sonnet) for natural language commands. The preference for terminal tools over traditional IDEs is noted, alongside tools like Cursor for navigation and remote access via Tailscale or SSH. Workload optimization strategies, such as reducing local hardware strain by offloading tasks to remote servers, and challenges with CPU limits on M1/M2 chips are also described.

Safety concerns with AI agents are emphasized, particularly the risks of "YOLO mode" (unrestricted operation) leading to unintended actions like misconfigured GitHub tickets. The discussion highlights the need for strict access controls, isolated environments, and safeguards to prevent destructive database interactions. Additional topics include the use of terminal-based tools for code review (e.g., Hunk, Delta), ASCII art creation with TermDraw, and experiments with AI agents interpreting diagrams. Open-source contributions, feedback aggregation tools like Modem, and the evolution of terminal-based interfaces (e.g., TUI libraries, mouse support, and GUI-like layouts) are also covered, alongside practical workflows for debugging, remote collaboration, and managing large-scale agent operations.

• What if you leveraged Tmux and Caffeinate for 24/7 agent workloads on a VPS?

  • Move: Set up a detached Tmux session on a VPS (e.g., Hetzner) with Caffeinate enabled to prevent sleep, running AI agents for batch tasks.
  • Why Now?: Your local machine is maxed out with multiple agents, and remote infrastructure offers scalability and stability.
  • Expected Upside: Continuous processing of non-coding tasks (e.g., log analysis, feedback synthesis) without local resource bottlenecks.

• What if you used Modem to curate AI-generated code for non-coding tasks?

  • Move: Deploy Modem to automate synthesizing customer feedback, managing ticket backlogs, and tracking user follow-ups via curated AI outputs.
  • Why Now?: Non-coding tasks are the "boring and slow" bottleneck, and AI can optimize this if paired with human oversight.
  • Expected Upside: Reduced manual effort in product management, faster prioritization of issues, and clearer user insights.

• What if you implemented JIRA/Linear-based approval blocks for agent actions?

  • Move: Configure tool approval blocks in JIRA or Linear to restrict AI agents from creating GitHub issues, updating tickets, or accessing private repos.
  • Why Now?: Uncontrolled agent actions (e.g., public GitHub tickets) risk violating workflows or security policies.
  • Expected Upside: Safer agent autonomy with clear boundaries, ensuring alignment with user intent and reducing accidental missteps.

• Implement Tmux and use VPS/droplet providers like Hetzner or DigitalOcean to manage continuous agent workflows, ensuring sessions can run unattended and be reattached across devices.
• Leverage Modem for automating non-coding tasks such as synthesizing customer feedback, managing ticket backlogs, and tracking user follow-ups, while curating AI-generated outputs through thorough reviews.
• Integrate JIRA or Linear into your workflow to enforce task prioritization and restrict agent actions to approved parameters, preventing uncontrolled execution (e.g., preventing agents from auto-creating GitHub issues).
• Ensure reliable infrastructure with tools like Caffeinate (to keep machines awake) and backup power solutions (e.g., portable batteries) to avoid agent stalling, paired with secure remote access via Tailscale for cross-device work.
• Apply strict access controls and avoid "YOLO mode" for AI agents by isolating sensitive data (e.g., GitHub, emails) and using VPS-hosted agents with limited scope to minimize risks of destructive actions or data breaches.