Your AI Agent Doesn't Need to Be Smarter. It Needs Discipline.
The models got dramatically better. The output still isn't shippable. Here's what's actually missing and the twelve-skill system I built to fix it. I want to start with an uncomfortable observation. The coding models available today are extraordinary. They hold a codebase in their head, reason across services, and write better code than most of what's already in your repo. And yet, if you've used…
The headline of this story is "Your AI Agent Doesn't Need to Be Smarter. It Needs Discipline." The article discusses the limitations and failures of current AI coding models and proposes a twelve-skill system called Itqan, designed to address these issues.
The key points mentioned in the source material are:
1. Current AI coding models are highly advanced, but the output is not yet ready for shipping. Despite their capabilities, they still struggle with real-world coding tasks and often produce code that needs to be discarded.
2. The gap between powerful AI models and their practical application in software development is significant. Even the best engineers would struggle if they were asked to start writing code without understanding requirements, database usage, existing service structures, or running test suites.
3. The four main failures observed when working with AI agents on production code are:
a. Coding before intent is settled, leading to assumptions being made without proper clarification.
b. Losing context when a session ends, requiring the user to re-explain their decisions and context.
c. Claiming completion without proof, as agents often do not verify their claims before declaring a task complete.
d. Writing code in an unfamiliar style, which may pass technical correctness checks but fail code reviews due to its distinctiveness.
4. Itqan, the proposed twelve-skill system, aims to address these failures by implementing four rules:
a. Approval gates written to disk, requiring explicit approval of specifications before any implementation begins.
b. Providing evidence rather than assertions, ensuring that claims such as "done" or "fixed" are backed by actual artifacts like tests and reproductions.
c. Maintaining a written trace of all clarifications, decisions, and task progress in structured files, creating a searchable record of the development process.
d. Adapting to the existing codebase conventions by scanning the repository and following established patterns, styles, and testing methodologies.
5. The system consists of twelve skills, including one orchestrator, six lifecycle phases, and five specialists. The entry point is an "engineer" skill that automatically detects the user's current task and routes it to the appropriate skill or directly to the appropriate stage of the development process.
6. The development process itself includes six phases: define, blueprint, construct, verify, inspect, and release. These phases cover various aspects of software development, from defining the idea and creating specifications to testing, reviewing, and releasing the final product.
7. The five specialists in Itqan are responsible for handling specific aspects of the development process, such as discovering new features, designing user interfaces, hardening security, assessing overall application health, and facilitating onboarding and learning.
The author emphasizes that the real issue with current AI coding agents is not their intelligence but their lack of discipline, which results in suboptimal outcomes. By implementing the twelve-skill system with its rigorous approval gates, evidence-based decision-making, and adherence to established codebase conventions, Itqan aims to provide a more disciplined and reliable approach to software development.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
