Andrew Ng made an observation this week: 90% of Agent projects stall at deployment — not because models aren't good enough, but because workflows lack guardrails. This week we noticed a Chinese open-source project, spec-superflow, tackling the same problem — just from a narrower angle, focused squarely on the "AI writes code" scenario.
What this is
spec-superflow is a suite of AI coding workflow plugins open-sourced by MageByte. At its core, it splits a single code change into 9 stages (explore, write spec, bridge contract, execute, debug, review, archive, merge), chained together by an 8-state state machine. It addresses two problems: first, AI jumping in before thinking clearly; second, beautifully written planning docs that the execution phase still ignores.
The core design is the "contract layer": it compresses the four planning-phase documents — proposal, spec, design, and tasks — into a single execution-contract.md. The execution phase only recognizes this contract, not chat history. Requirements drift triggers mandatory rollback; no approval, no touching code. The whole pipeline sets up three human gates, the most critical being "contract approval" — without a nod, the AI doesn't even qualify to start execution.
Industry view
Supporters argue this is the marker of AI coding tools graduating from "toy" to "engineered." The planning-execution disconnect is a common pain point across tools like Claude Code and Cursor; spec-superflow uses hard constraints rather than soft prompts to lock down the workflow — a notably more advanced approach.
But the dissent is clear. First, the learning cost of 9 Skills plus an 8-state state machine is extremely high; small teams and individual developers may well find "time spent configuring the workflow exceeds time spent coding." Second, the state machine's hard interception means low fault tolerance — a single requirement change can trigger a rollback to exploring, and overall efficiency may not beat manual review. Third, there's no public data proving its stability in large real-world projects; the demonstrated "add RBAC" is still a typical simplified scenario, still far from production-environment scrutiny. Fourth, the sustainability of the open-source project is in question: is MageByte led by an individual or a company? Whether subsequent maintenance and community governance are transparent determines whether it can enter enterprise procurement shortlists.
Impact on regular people
For enterprise IT: If you're already using AI for coding and have been burned by "discovering the direction is wrong only after the changes are made," tools with state machines and human gates are worth evaluating — but we recommend running them on non-critical projects for 1–2 months before drawing conclusions.
For individual careers: Developers don't need to panic about being replaced, but they do need to add a new skill — writing requirements as verifiable specs (SHALL/MUST + Scenario). This is becoming fundamental literacy in the era of AI collaboration.
For the consumer market: No direct short-term impact on consumers, but it will indirectly affect the iteration speed and quality stability of all software products — especially the small SaaS tools that rely on AI to rapidly prototype.