WILLOSCAR/research-units-pipeline-skills

Research pipelines as semantic execution units: each skill declares inputs/outputs, acceptance criteria, and guardrails. Evidence-first methodology prevents hollow writing through structured intermediate artifacts.

What it solves

Research Harness provides infrastructure to make agent-assisted research inspectable, resumable, and honest. It solves the problem of "black box" research agents that produce a final answer (like a PDF) but leave no trail of evidence, making it difficult to verify which sources support specific claims or to resume work after a failure.

How it works

The system turns a research goal into a file-first, recoverable "Run." It organizes research into focused Skills and explicit Workflows. The process follows a linear progression: Goal $\rightarrow$ Run $\rightarrow$ Evidence $\rightarrow$ Artifact.

Key mechanisms include:

  • Pinned Contracts: The system snapshots the pipeline and hashes its components to ensure a Run cannot silently continue if the underlying rules or skills change.
  • Evidence-Backed Completion: Success is not just a "DONE" status; it requires agreement between attempts, required outputs, artifact hashes, and workflow checks.
  • Bounded Diagnosis: When a contract fails, the system distinguishes between observable defects and the specific repair surface, allowing for targeted improvements rather than rewriting the entire process.
  • Human Checkpoints: Users can approve specific artifacts (like an outline or protocol) which are bound to hashes, meaning any change to the approved item revokes the authorization.

Who it’s for

Researchers and developers using AI agents to conduct literature reviews, synthesize studies, or write reports who need a rigorous, verifiable trail of evidence and the ability to recover from failures.

Highlights

  • Diverse Workflow Templates: Includes pre-defined workflows for research briefs, paper reviews, evidence reviews, and LaTeX/PDF survey generation.
  • File-First Architecture: Maintains a workspace containing the goal, unit state, decisions, and a full provenance trail in a .harness folder.
  • Strict Integrity Checks: Implements execution integrity and contract acceptance layers to ensure deliverables meet observable workflow checks.
  • Automated Diagnosis: Includes tools to diagnose failures and point them back to the smallest repair surface.

Related

  • Project
    ApodexAI/AgentHarness

    An evaluation harness for testing AI agents on deep-research benchmarks using a standardized ReAct setup.

  • Project
    appautomaton/latex-arxiv-SKILL

    An agentic harness for writing verified AI/ML review papers in LaTeX, featuring a gated, issue-driven workflow that prevents citation hallucinations.

  • Project
    OpenNSWM-Lab/FAROS

    A multi-agent collaborative system for AI scientists that automates the research loop from ideation and literature retrieval to code execution, paper writing, and auditable peer review.

  • Dispatch
    Beyond the Model: The Rise of Harness Engineering for AI Agents

    Explore the concept of Harness Engineering, a systematic approach to making AI coding agents reliable through environment design, state management, and explicit constraints.

  • Project
    greyhaven-ai/autocontext

    A recursive self-improvement harness that automates the process of running, evaluating, and refining AI agent strategies to create durable playbooks and datasets.