Quick disclosure upfront: this is my own project. Not dropping a link to farm clicks — the context matters for the post.
I've started and abandoned four agent projects in the last year. Every single time I stalled at the same place: orchestration. I had agent ideas, prompts, even decent tool-use loops. What I didn't have was a backbone — something that said 'agent 3 finishes, agent 4 picks up the artifact, agent 7 audits it, and the whole thing doesn't silently die at 3am.'
I kept reaching for LangGraph, CrewAI, AutoGen, rolling my own with Redis queues. Every path turned into two weeks of infra before I wrote a single line of actual agent logic. Then I'd lose momentum and shelve it.
What finally worked for me was brutally boring: 28 markdown files in an Obsidian vault defining agent roles, handoffs, and artifacts, plus 3 Claude Code harness YAMLs that actually execute the loop. No framework, no vector DB, no orchestrator service. The vault IS the orchestrator — Claude Code reads it, runs the agents, writes artifacts back. 10 agents, deterministic handoffs, observable because every step is a file you can open.
This is the same fleet running mindsparkstack.com right now. Content pipeline, research, QA, publishing — all of it.
I packaged the exact vault + harness as Fleet Architect. $50 one-time, no subscription, no SaaS wrapper. You clone the vault, point Claude Code at it, and have a working 10-agent fleet in an afternoon. It's not magic — it's the scaffolding I wish I'd had on attempt one.
Happy to answer specifics in comments — agent roles, how the handoffs work, why Obsidian instead of a 'real' orchestrator, where it breaks. Genuinely curious if others solved this differently, because I spent too long assuming the answer had to be complicated.