rpamis/comet

Comet: agent skill harness for turning ideas into evaluated workflows

What it solves

Comet addresses the instability and lack of structure in long-running AI coding tasks. It prevents agents from "forgetting" progress or guessing state after interruptions and replaces intuition-based skill improvement with a scientific evaluation framework. It specifically solves the problem of coordinating complex requirements that are too large for a single agent session by splitting them into manageable, dependency-aware child tasks.

How it works

Comet provides a resumable runtime and a skill platform that integrates with AI coding tools. It offers two primary workflows:

  • Native: Designed for strong models, allowing them to autonomously plan and verify while Comet handles state checks and resumable archiving.
  • Classic: A constrained five-phase methodology (OpenSpec + Superpowers) using a state machine to ensure strict adherence to a specific process.

For complex goals, it uses Supervisor Changes to create a Directed Acyclic Graph (DAG) of child changes, executing them in isolated worktrees via multiple agents before integrating them. It also includes a memory system for personal preferences and project knowledge, and an evaluation platform (comet eval) that uses Rubrics and Pass@k metrics to measure skill performance.

Who it’s for

It is built for developers using AI coding agents (like Claude Code or Cursor) who need to manage complex, multi-step software engineering tasks that require strict verification, resumability, and scientific optimization of their agentic workflows.

Highlights

  • Resumable Workflows: Task state is stored in the project, allowing work to resume exactly where it left off after interruptions.
  • Supervisor Changes: Coordinates multiple agents to implement complex requirements in parallel across isolated worktrees.
  • Dual Workflow Modes: Offers both a flexible "Native" mode for autonomous models and a structured "Classic" mode for high-constraint tasks.
  • Skill Platform: Tools to author, package, and distribute reusable coding skills across 37 different AI coding platforms.
  • Scientific Evaluation: Integrates with LangSmith and LangFuse to evaluate skills using Rubric, Pass@k, and Pass^k scoring.

Related

  • Project
    dongshuyan/compass-skills

    COMPASS Skills is an alpha‑stage, modular toolkit of ten markdown‑based AI agent skills (task clarification, task graph, pause‑resume, hand‑off prompt, user profile, skill‑builder/upgrader, clear communication, academic humanizer, interview assessment) that let LLM‑driven agents persist context across conversations, align with user preferences, and safely evolve from real run evidence. Install via `npx skills add …` or copy the skill folders into a harness’s skill directory.

  • Project
    hellowind777/helloagents

    A workflow layer for AI coding CLIs that adds specialized skills, project knowledge bases, and verification gates to ensure agents complete tasks reliably.

  • Project
    kangarooking/kangarooking-skills

    A collection of structured, open-source Agent Skills compatible with the Agent Skills standard, providing reusable workflows for content creation, image generation, and AI project engineering.

  • Project
    deepagent-ltd/deepagent-code

    DeepAgent Code (DAC) is a desktop‑plus‑terminal workspace that lets you use an LLM as a coding assistant for multi‑step tasks. It offers persistent sessions, inspectable project memory, sub‑agents for independent jobs, a managed browser, and optional instant‑messaging integration, all configurable with any compatible LLM provider. Licensed under AGPL‑3.0.

  • Project
    davidondrej/skills

    A curated set of prompt‑based skill definitions for AI coding, research, and ops agents, organized by workflow type and ready to be loaded into Cloudroom or similar agent runtimes.