# Turning Tokens Into Passable React --- ## <u>OR:</u> How I Raised a Small Army of Bots to Clean Up Our Messes and It Mostly Worked, Kind Of --- ## If you want to skip to the end go to https://staging-next.voze.com --- # Welcome to The Preamble ---  ---  ---  --- # What Changed? --- ## Opus 4.6 and Subagents and Agent Teams --- ## Maybe More Important: Building Virtual Engineering Organizations --- <!-- .slide: class="center-large" --->  --- # How I Set It Up --- ``` I want to use a virtual team workflow to [DESCRIBE GOAL]. Follow this process: ### Phase 1: Architect Plans Launch a [DESCRIBE ARCHITECT TYPE] architect agent to thoroughly explore the codebase, then create a sprint plan that: - Breaks work into N parallel engineer assignments (typically 3-5) - Ensures each engineer's scope is independent (no file conflicts between engineers) - Assigns roughly equal work per engineer - Specifies exact files each engineer should read/modify - Requires each engineer to write tests and run the full test suite before completing - Write the plan to .sprint/SPRINT_PLAN.md for persistence ``` --- ``` ### Phase 2: Engineers Execute in Parallel Launch all engineer agents, no more than [MAX AGENT COUNT], simultaneously. Each gets a FOCUSED prompt with: - Only the context they need (specific files listed, not the whole codebase) - Clear scope boundaries ("Do NOT modify files outside your scope") - Explicit deliverables (code changes + tests) - Requirement to run the full test suite before completing - Write status to .sprint/ENGINEER{N}\_STATUS.md when done ``` --- ``` ### Phase 3: Cross-Reviews in Parallel Each engineer reviews another's work (circular: E1→E2, E2→E3, ..., EN→E1). Use Sonnet for reviews to save cost. Each reviewer: - Follows a review checklist (test quality, patterns, consistency) - FIXES issues directly rather than just reporting them - Runs affected tests after fixes - If they make fixes, those then need to be reviewed by another engineer agent before they are considered done - Writes review to .sprint/REVIEW_E{N}\_OF_E{M}.md ``` --- ``` ### Phase 4: QA Validation Single QA agent runs: - Full test suite - Type checking / linting - Feature verification against the sprint plan - Updates the project review/status document - Writes final report to .sprint/QA_REPORT.md ``` --- ``` ### Rules: - All sprint artifacts go in .sprint/ directory, cleaned up between sprints - Keep engineer prompts small and focused — list only their specific files - Update the sprint plan status after each phase completes - If engineers must touch the same file, ensure they modify completely different sections (different line ranges) and document the coordination - Engineers should use backward-compatible changes when extending shared code (new optional params with defaults so existing consumers aren't broken) ``` --- # How It Went --- ## Burned Through So Many Tokens --- ## Go Check Out staging-next.voze.com --- ## Creating UI Took More Prompting Than You Would Think ---  --- ## Sometimes Subagents Just Die --- ## The Context Spiral of Death --- ## Have Them Write Things Down All The Time --- # The Upshot --- # ?s ---