Employees, AI, and AI Employees (slobodan.me)

Two founders documented a 36‑day, five‑month experiment trying to build an “AI cofounder” (CofounderGPT) and concluded it failed — not for lack of clever demos but because current LLMs lack sustained agency, legal standing, emotions, and true responsibility. The piece argues that while models and tools (GPT-5 Pro, Claude Opus, Cursor, Claude Code, Codex) have dramatically improved coding, reasoning, and tool integration, key technical limits remain — context window size, robust multi‑step reasoning, reliable tool chaining, and verifiable decision‑making — that prevent an AI from being a full cofounder who can accept consequences or steer long‑term strategy. For the AI/ML community this frames a practical roadmap: most software engineering activities will be augmented (faster research, scaffolding, tests, API schemas) or commoditized (routine code, payloads, CI pipelines), but higher‑order work — intent/prioritization, domain modeling, security tradeoffs, incident response, and long‑term architecture — will stay human‑centric for the foreseeable future. Using a Wardley‑style evolution map, the author predicts a shift from “employee + AI” (custom tool use) toward more standardized AI employees over time, but warns that legal, ethical, and reliability questions must be solved before true automation of knowledge work is possible.