Orchestrating Multi-Agent Workflows in 2026
Evolved from hand-coding to spec-driven agent orchestration, multitasking 2-4 agents via git worktrees in Superset, blending product/marketing tasks to overcome single-agent bottlenecks.
Workflow Evolution Unlocked Spec-Driven Multitasking
Brian Casel traces his shift from manual crafting in 2023—designing, coding, and writing by hand—to AI-enhanced work in 2024, collaborative spec-driven development in 2025, and full agent orchestration in 2026. The key insight: 90% of his time now shapes ambitious specs, with only 10% reviewing outputs. Single-agent workflows bottlenecked as specs grew complex, pushing him to multitask 2-4 agents across features or codebases simultaneously. This wasn't just better models (Claude 3.5+ or harnesses like Claude Code); it stemmed from reallocating creative energy to product thinking. "Spec writing is a practice just like any other craft... once those specs became larger and more ambitious, working with just one agent at a time started to feel like a bottleneck."
He rejected sequential workflows for parallel execution using git worktrees, avoiding overwrites by isolating agent tasks in repo copies. Before: IDE + terminal for one task. After: Sidebar-driven interfaces flipping between worktrees/projects. Result: Faster iteration, with commits/PRs/merges/deployments streamlined (e.g., Cmd+Shift+P for PR setup, auto-deploy on main merge, optional staging).
Tool Selection Prioritizes Native Claude Code and Worktree Fluidity
Casel evaluated Cursor v3 (agent sidebar, multi-repo but no native Claude Code CLI), Claude Desktop (redesigned code tab, plan views, but inconsistent file access/skills/keyboard shortcuts), Conductor (Claude wrapper, auto-worktree creation but secondary terminal/Claude Code support), and Superset (current daily driver). Tradeoffs: Cursor/Conductor wrap Claude (losing native CLI sounds/icons); Claude Desktop lacks full file browser; Conductor hides local branch easily.
Superset wins for native Claude Code execution, default local branch access, Cmd+N for instant worktree spins (e.g., "add a joke to the footer of homepage"), full filetree visibility, and terminal tabs for git pulls/commits. Layout convergence across tools (left sidebar repos/worktrees, center chat, right files/changes) suits non-coding oversight. He still jumps tools (Superset v2 upcoming) but focuses sidebar multitasking. "I like how in superset... at any given time I'm typically flipping between different work trees within one project and sometimes I'm even working on additional projects and I can flip between them uh just like that."
For research-heavy specs/reviews, he recommends Consensus MCP server: queries 200M+ peer-reviewed papers with citable line-by-line sources, integrable via Claude/Cursor/etc. Superior to web-scraped noise for high-stakes (healthcare/fintech) decisions. Use cases: Validate UX/security tradeoffs pre-spec; audit PR decisions post-build.
Blending Product and Marketing via Reusable Agent Skills
Casel builds internal tools like Spark Drop (content idea pipeline/newsletter scheduler) and Brain Down (Dropbox-integrated markdown editor/sharer), then extracts public assets in parallel worktrees. Reusable skills (e.g., app-marketing-page, starter-kit-generator, plan-videos-from-build) analyze codebases, interview him, generate PRDs/tech stacks/prompts, design/implement pages, outline videos.
Process: Main skill.md orchestrates phases; steps/ folder details instructions. For marketing pages: Analyze repo → interview messaging → frontend design → implement one-pager (e.g., braindown.app/sparkdrop.co homepages without SaaS sales, just explanations/starter kit teasers). Starter kits for Pro members include his specs/prompts/PRDs/video guides for customization. Video skill pulls code/conversations for YouTube outlines.
This fuses workflows: Product builds and marketing (pages/assets/videos) run side-by-side, even same codebase. Tradeoff: More upfront skill crafting, but scales operations. "Product and marketing used to be two totally separate workflows and now they run in parallel sometimes inside the same codebase."
Mobile Enables Autonomous Overnight Agent Tasks
Mobile Claude Code (iOS app) addictively extends multitasking: Queue big tasks (codebase analysis/research/spec churn) pre-dinner/bedtime/morning routine, optimizing wait times. Patterns: Evening office handoff → sleep work → morning review/queue next. Blends with exercise/coffee. Enables true autonomy as agents handle long-running tasks untethered from desktop.
Key Takeaways
- Allocate 90% effort to spec shaping; review 10%—push creative energy there to justify multi-agent scale.
- Use git worktrees (via Superset Cmd+N) for safe parallel agent tasks across 2-4 features/repos.
- Build reusable agent skills (main.md + steps/) for ops like marketing pages/starter kits/video plans to blend product/marketing.
- Prioritize native CLI tools (Superset > wrappers) for familiar DX; evaluate sidebar layouts for filetree/chat fluidity.
- Integrate Consensus MCP for research-backed specs/PR reviews in high-stakes builds.
- Leverage mobile for autonomous tasks: Queue overnight to reclaim personal time.
- Evolve sequentially: Hand-code → AI enhance → Collaborate → Orchestrate multitask.
- Extract assets (starters/videos) from builds to fuel marketing/content pipelines.
- Default to local branch workspaces; merge via PRs for deploy safety.