Video brief · S01E04

Working with AI Harnesses

Brief in progress — review the script and demo before recording.

In this video, we need to show that the harness is not just about what happens right now and how to work with 5 agents at once, but also about how we can push it further with MCPs, the CLI, and personal, general-purpose tools. But first and foremost, we need to show how to make our setup harness-agnostic.

  • Principles: things everyone should know about harnesses: core features, providers, flexibility, cloud and background workflows, integrations, extensions, and programmatic setup
  • Agnostic Setup: the motivation for remaining harness-agnostic, skills, a knowledge base, extensions, surrounding tooling, and common and personalized workflows
  • Personalization: how to set up a harness tailored to our work and projects, and how to keep improving it by scanning history, extracting skills, and building reusable tools
  • Programmatic Workflow: a harness can manage itself, including by spawning instances. We need to show how a single agent can manage others and how this approach can support workflows, better handoffs, and background work.
  • Teamwork: workflows for shared setups and specs for teams, as well as code review, quality checks, communication, etc.
  • Background Tasks: tooling for background tasks (basic tools like GitHub Actions or Cursor agents) and bots for messaging apps. A Grok bot and/or Claude schedules may also be good examples of automation beyond the codebase, as well as in the work around it.

We may want to showcase scenarios such as:

  1. This one should be simple: completing the same task with 3 popular harnesses to demonstrate harness-agnostic delivery.
  2. Showcasing two examples of background tasks that improve quality, and one that improves the system itself by scanning history from a given day.