Haft for AI Coding Assistants — A Decision Engineering Runbook

2026-05-04 · Source: Artificial Intelligence in Plain English - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, medium

Summary

Haft is a decision-engineering governor for AI coding assistants, designed to sit between user intent and agent execution, implementing a five-mode cognitive cycle. It addresses common failure modes like lost rationales and un-revisited assumptions by enforcing structured decision-making. Haft uses seven typed Model Context Protocol (MCP) tools, such as `haft_note` for micro-decisions and `haft_reason` for deeper analysis, which map to framing, exploration, commitment, execution, and lifecycle hygiene. The system installs via a one-liner `curl` command and `haft init`, supporting Claude Code and Codex, with experimental paths for Cursor and Gemini. Haft stores decisions locally in SQLite, projects them as Git-trackable markdown, and uses `haft sync` for team reconciliation. Its v7 architecture prioritizes a parseable `ProjectSpecificationSet` before broad harness execution, ensuring agents operate on well-defined systems.

Key takeaway

For AI Engineers and MLOps Engineers seeking to formalize decision-making with AI coding assistants, Haft provides a structured framework to prevent lost rationales and ensure verifiable outcomes. You should integrate Haft to enforce a cognitive cycle that demands explicit framing, alternative exploration, and recorded decision contracts, moving beyond ad-hoc chat interactions. This approach helps maintain architectural integrity and provides auditable evidence for future system changes, reducing technical debt from underspecified agent outputs.

Key insights

Haft governs AI agent execution by enforcing structured decision-making and preserving rationales through a five-mode cognitive cycle.

Principles

• Enforce distinct alternatives before commitment.
• Degrade trust scores when evidence ages.
• Prioritize system specification before broad execution.

Method

Haft employs a five-mode cognitive cycle (Understand, Explore, Choose, Execute, Verify) using MCP tools like `/h-frame`, `/h-explore`, `/h-compare`, `/h-decide`, and `/h-verify`, with `/h-note` for reversible micro-decisions and `/h-reason` for automated depth.

In practice

• Use `/h-reason` to force agents to surface distinct options.
• Track decisions in Git as markdown projections.
• Run `haft spec check` for local structural validation.

Topics

• Haft
• AI Coding Assistants
• Decision Engineering
• Model Context Protocol
• Cognitive Cycle

Code references

• m0n0x41d/haft
• ailev/FPF
• m0n0x41d/quint-code

Best for: AI Engineer, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Coding agents are giving everyone decision fatigue — Coding agents shift software development bottlenecks from code generation to human judgment, intensifying work and causing decision fatigue.
• LAI #111: The Craft Layer of AI -Voice, Speed, and Real-World Interfaces — Effective AI control spans output quality, system performance, and secure enterprise integration for real-world transactions.
• MiroThinker-1.7 & H1: Towards Heavy-Duty Research Agents via Verification — MiroThinker-H1 enhances research agents with integrated verification for reliable, multi-step complex reasoning.
• From Dashboards to Decisions: How We’re Building AI Agents at RaySuite AI — AI agents transform user interaction by providing intelligent, contextual answers and taking action over complex systems.
• Presentation: State of Play: AI Coding Assistants — AI coding assistants are evolving towards greater autonomy, necessitating advanced context engineering and robust safety harnesses.
• AIE Europe Day 2: ft Google Deepmind, Anthropic, Cursor, Factory, Linear, HF, Cerebras & more — AI engineering is rapidly evolving, demanding new approaches to agent orchestration, code quality, and human-AI collaboration.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence in Plain English - Medium.