# Lifestyles of the AI-Native — Nick Nisi & Zack Proser, WorkOS

## Executive summary

This workshop details advanced workflows for operating 'AI-Native' engineering teams, moving beyond single-session agent interaction to managing 'fleets' of autonomous agents. Key concepts include defining measurable completion criteria using 'Goals,' setting up repeatable tasks with 'Loops,' enforcing mandatory steps via 'Hooks,' and parallelizing development using 'Git worktrees.' The focus is on building robust, autonomous systems that minimize human babysitting while preserving human judgment for final review.

## Key takeaways

- Shift from Single Session to Agent Fleet: Engineers should operate a 'fleet of agents' rather than babysitting a single session, allowing for concurrent work across multiple tasks and projects.
- Goals vs. Loops: A 'Goal' has a clear, measurable stopping condition (e.g., 'refactor this until all tests pass'). A 'Loop' repeats a task until manually canceled or a timer expires.
- Autonomous Workflow Primitives: The system uses 'Hooks' (mandatory steps), 'Goals' (measurable termination), and 'Loops' (repetition) to ensure tasks complete fully, preventing agents from 'punching out' early.
- Concurrent Development with Worktrees: Using 'Git worktrees' allows agents to run in parallel on isolated, safe copies of the repository, compressing large development blocks into shorter timeframes.

## Technical details

- Voice Coding: Using local dictation tools like Handy (which runs on-device) allows engineers to operate at the 'speed of thought,' significantly accelerating the process of generating code, file names, and technical commands.
- Agent Orchestration (Fleet): Tools like 'Fleet' help manage and track multiple concurrent cloud sessions (e.g., 12 agents running simultaneously) by providing a centralized summary and status indicator.
- Verification Gates and Hooks: Hooks enforce mandatory steps (e.g., running linting, type checks, and tests) before an agent can proceed, preventing the system from lying about completion. Adversarial review (using a second model like Codeex) is recommended to find issues locally.
- Scheduled Tasks: The 'schedule' command allows automating tedious, recurring tasks (e.g., weekly reports, dependency updates) on a durable, repeatable basis, acting like a cron job for AI workflows.
- Model Context Management: To prevent context loss in long sessions, advanced techniques involve breaking work into small, atomic pieces and using state machines (e.g., in TypeScript) to force the agent through defined steps.

## Practical implications

- Implement structured agent workflows using Goals and Loops to ensure tasks reach a verifiable state of 'done,' rather than simply appearing complete.
- Adopt local, on-device dictation tools to maximize development speed and maintain data privacy during coding sessions.
- Utilize Git worktrees to parallelize development efforts, allowing multiple agents to work on isolated code branches concurrently.
- Automate recurring operational tasks (reporting, dependency checks) using scheduled tasks to minimize manual overhead.

## Topics

AI Agents, Workflow Automation, Software Development Lifecycle, LLM Orchestration, DevOps, WorkOS workshop repository, Handy, Fleet, Claude Code / Claude

Source: https://www.youtube.com/watch?v=AgBAOXVxt4A