954 completed tasks out of 1,002 in two months—this isn't a post-funding report from a hyped startup, but the output of a single script running on 600 lines of code. While the enterprise sector shells out for over-engineered platforms and debates AI ethics, real margins are being found in architectural minimalism. The experience of SKUmind—an analytics service for Ozon marketplace sellers—proves that the primary bottleneck in automation isn't the "stupidity" of models, but the human serving as a "living router."
The problem faced by the project's founder is typical for any scaling business: he became the bottleneck. Even with 12 open Claude Code tabs, where each AI played a specific role from tech lead to media buyer, he spent his days summarizing context. The tech lead produced a migration—he had to carry it to the tester. The analyst refined a spec—he had to run it to the developer. In this configuration, system performance is capped by the speed at which a Homo Sapiens can copy-paste text between windows. It creates an illusion of productivity, but in reality, it's a hard ceiling on growth.
The Failure of Trendy Concepts
Attempts to solve coordination through popular "workarounds" predictably failed. Messing with environment variables crashed against the limitations of VS Code, which struggles to isolate contexts for different panels. Using a Telegram bot as a transport layer looked elegant only in presentations; it solved delivery but ignored structure. Without a rigid protocol, a group chat instantly turns into a digital swamp where "everything is written, but nothing can be found." A third path—injecting rules via a CLAUDE.md config file—proved architecturally messy, cluttering product code with operational noise.
Information flowed through me. I wasn't an architect; I was a packet router. And a router with bad memory at that: by evening, I couldn't remember what I’d handed off or what was still pending, and sometimes I lost track of what I’d even done all day.
This "trial by fire" phase is critical. It proves that autonomy doesn't require complex AI orchestrators or expensive SaaS subscriptions. It requires a simple coordination bus built on Python’s standard library that ignores marketing noise and simply moves data. By stripping away unnecessary entities, the team discovered a mechanic where oversight no longer requires human intervention.
Blackboard Architecture vs. Corporate Chaos
The working solution turned out to be the Blackboard pattern—a "shared board" implemented through a standard markdown file. This is a classic network topology where any participant can pass work to any other without hitting a central hub. In SKUmind’s system, each of the 12 Claude agents has its own section in the file with strict task statuses: open or done. No databases, no servers—the file lives locally, eliminating conflicts during parallel writes from different windows.
The key element is the rigid demarcation of roles (boundaries) in the configuration. The tech lead handles architecture and reviews but stays out of the specs. The analyst defines what to build but doesn't touch the code. This segmentation allows agents to act autonomously, guided by data structure rather than "improved prompts." In this setup, the markdown file becomes the single source of truth, replacing endless meetings and Jira tickets. A median of 47 minutes per task is a death sentence for traditional departments, where a single "handoff" approval can drag on for days.
The economics of this architecture are transparent: it saves on token costs by cutting out irrelevant context and, more importantly, gives the founder his life back. When a system handles 1,002 task handoffs in 60 days without human help, the question of hiring new staff disappears. This radical simplification exposes the inefficiency of enterprise solutions that stack layers of abstraction where a text file and a few hundred lines of code would suffice. Large companies will either have to embrace this minimalism or continue burning budgets on meaningless tickets until they are disrupted by those who know how to write scripts.