ARIS: A Markdown-Only Research Workflow That Lets LLM Agents Audit Each Other

The Problem: Unreliable Autoresearch and the Need for Audit

Autonomous ML research with LLM agents is appealing, but it has a credibility problem. Agents can produce plausible-sounding results that do not survive scrutiny. The README for ARIS (Auto-Research-In-Sleep) directly addresses this: it positions the project as a way to run research loops where other models review each step. The core issue is that a single model judging its own work tends to miss its own errors. ARIS attacks that by enforcing cross-model review loops. The intended audience is researchers and engineers who already use agentic coding tools like Claude Code, Codex, or OpenClaw and want a structured workflow for idea discovery, experiment design, and execution. It is not for people who want a black-box platform; it is a methodology delivered as Markdown skills that you can read and modify.

How It Works: Skills, Review Loops, and a Research Wiki

ARIS is not a framework or a library. It is a collection of Markdown files, each defining a skill. The README describes it as a 'skill-based workflow' that works with several agents. The mechanism is straightforward: you give the agent a research task, and the skill files guide it through steps: ideation, experiment design, execution, and review. The review loop is the key. The README mentions that ARIS keeps a 'research-wiki for memory' and has 'other models check every frame.' That memory is important because long-running research tasks often lose context. The wiki gives the agent a persistent store of findings and decisions. The cross-model aspect means a different model (or a separate instance) reviews the work, reducing self-consistency bias. The README also points to a companion project, Anti-Autoresearch, which audits the output for '61 signals' of fabrication and AI-style impressions. That is a separate tool, but it shows the philosophy: verify, do not trust.

Getting Started: Real Commands and Configuration

The README gives concrete entry points. For the standalone experience, you can use the ARIS-Code CLI, but the simplest path is to use the skills directly in your existing agent. For Claude Code, you would install the skills into your project. The README does not give a one-line install for the base version, but it does for the DeepSeek Harness plugin: `dsh plugin --profile web add dsh-aris`. That command fetches the plugin from npm, but it requires `pnpm` on your PATH. For the base workflow, you need to clone the repository and copy the `skills/` directory into your agent's skill folder. The README references a `CLAUDE.md` or `AGENTS.md` file where you can add a block from the HERO project to bound over-defensive behavior. The ARIS-Monitor is a separate utility: `cd aris-monitor && ./run.sh` launches a small macOS widget that shows when a session waits for approval. That is a nice touch, but it is optional.

The DeepSeek Harness Release: A Plugin That Installs Itself

The latest release, `dsh-aris-v0.1.0`, is specifically for the DeepSeek Harness. The README says it installs as one plugin and that all 82 skills are unchanged, with Codex still acting as the independent reviewer. That is a significant detail: the plugin provides a packaged experience, but it does not alter the core methodology. The setup is simple: one command, no separate install step, but you must have `pnpm` available. This release is useful if you are already on DeepSeek Harness, but it also signals that ARIS is actively evolving. The main branch has releases like `v0.4.24` from August 2026, so the versioning is active. If you are not on DeepSeek Harness, you do not need this release, but it shows that the project is not stagnant.

Limitations and Failure Modes

The biggest limitation is that ARIS is only as good as the skills you use and the models you run. The README claims to catch unreliable autoresearch, but that is a strong claim. The cross-model review loop depends on having access to multiple models, and if you only have one, the loop collapses. Also, the research-wiki memory is a plain Markdown file, so it can grow unwieldy on long projects. The README mentions that long stories break when the model forgets earlier details, which is exactly the failure mode ARIS tries to mitigate, but it does not eliminate it. Another limitation is that ARIS is not a framework, so there is no automated scheduler or resource management. You are still responsible for running the agents and monitoring them. The ARIS-Monitor widget is a band-aid, not a full solution. If you need a hands-off system that runs overnight without supervision, ARIS may not be enough.

Alternatives: ARIS vs. Full Agentic Research Platforms

The natural alternative is a full agentic research platform like AutoGPT or a dedicated ML experiment runner. These tools provide a UI, task queue, and sometimes a built-in model. ARIS takes the opposite approach: it gives you Markdown skills that you slot into your existing agent. The difference is in the level of integration. A platform like AutoGPT abstracts away the agent, but it also locks you into its design. ARIS is transparent and portable, but it requires you to do more assembly. Another alternative is to write your own prompts and review loops from scratch. That gives you total control, but you lose the curated 82 skills and the anti-autoresearch audit. ARIS sits in the middle: it gives you a structured methodology without a heavy runtime. The trade-off is that you must trust the skill definitions and the audit signals, which are not independently verified.

Maintenance, Upgrade Cost, and License

ARIS is MIT-licensed, so you can use, modify, and redistribute it freely, but the license does not come with warranty or support. The project is actively maintained, with releases in August 2026. The maintenance cost is low because the core is just Markdown files. Upgrading is a matter of pulling the latest version and checking the skill definitions for changes. However, the README points to several companion projects (Anti-Autoresearch, HERO, ARIS-Movie-Director), and that ecosystem adds complexity. You may need to track updates across multiple repositories. The DeepSeek Harness plugin has its own branch, so if you use that, you need to follow that branch separately. The upgrade cost is not high, but it is not zero. You should read the release notes for each version to see if any skill behavior changed.

Editorial conclusion

Adopt ARIS if you are an ML researcher or engineer who already uses an LLM agent like Claude Code or Codex and want a structured, repeatable research loop without committing to a heavyweight platform. The Markdown-only design means you can inspect every skill and adapt it to your own workflow. Skip it if you need a turnkey solution or prefer a single integrated tool. Before adopting, verify that the cross-model review loop actually catches the failure modes that matter to you: the README claims audits and anti-autoresearch signals, but you should test them on your own experiments, especially the integrity checks that distinguish reported gains from real ones. Also check the version you plan to use, since the latest release is on the dsh-aris branch and the main branch may lag. The project is under active development, so pin a release and read the AGENT_GUIDE.md before running it unattended.

Official sources

1. Official README
2. Project repository
3. Release notes