The Good, The Bad, And The Ugly: Workflows aren't all you need
Details
LOCATION CHANGE: Normally we meet at the Clifton Branch Library, but we'll be meeting at the Hyde Park Library due to availability issues.
Speaker: Sean Thimons
The Good, The Bad, and The Ugly: Workflows Are Not All You Need
AI coding assistants are evolving from pair programmers into autonomous workflow engines. With custom slash commands, reusable skill templates, and multi-agent orchestration, a solo developer can now execute structured software lifecycles — planning, implementation, testing, review, and deployment — at a pace that was previously impossible. But what happens when the workflows work too well?
Over 256 sessions and 91 commits in 26 days, I used Claude Code's agentic capabilities to build and operate a structured development framework called GSD (Get Shit Done) — a phase-based workflow system with discussion, planning, execution, and verification gates, powered by parallel subagent orchestration. The results were impressive (on paper): 80% fully-achieved outcomes, 56,000+ lines shipped, and an autonomous content pipeline generating 20+ blog posts with minimal intervention.
But the story doesn't end there. Beneath the productivity numbers lie recurring failure modes that no amount of workflow engineering could prevent: an AI that confidently writes inverted narratives about data it never verified, a Windows environment that the agent forgets about mid-pipeline, and — most dramatically — six parallel milestones developed by independent agent sessions that converged into a 135-file, 28,000-line integration nightmare requiring reverts, orphaned PRs, a dedicated bug bash, and four follow-up hotfixes.
This talk examines what structured AI workflows actually buy you (repeatability, parallelism, institutional knowledge), what they don't (semantic correctness, cross-workflow coordination, judgment under ambiguity), and the uncomfortable maintenance burden of keeping the workflow machine itself running. The central argument: workflows solve the process problem but not the judgment problem, and parallel autonomous workflows without a convergence strategy are a divergence engine that builds technical debt at the speed of AI.