OpenAI’s recent push into agentic coding has sparked debate. The company touts its Symphony orchestrator as a proof that a specification file can spawn a full project. Yet the spec, named SPEC.md, reads more like pseudocode than a clean interface, blurring the line between design and implementation.

Critics argue that the spec’s heavy use of prose dumps and embedded code fragments defeats the claim that specifications are simpler than code. When the document lists database schemas, concurrency limits, and retry logic in plain text, it essentially mirrors the logic the agent must later generate, making the distinction moot.

An attempt to let Claude Code build Symphony in Haskell illustrates the limits of current agentic tools. The resulting repository, Gabriella439/symphony-haskell, contains incomplete or incorrect code, underscoring that even advanced LLMs struggle to translate a pseudo‑spec into a working system without human guidance today effort.

These findings suggest that agentic coding remains a niche experiment rather than a wholesale replacement for human developers. Until specifications can be expressed in a truly abstract, machine‑readable format that eliminates the need for embedded code, the promise of fully automated code generation will stay on the horizon.