OpenResearch: Reproducible Agent Research

Overview

OpenResearch frames agent-assisted research as a reproducibility and attribution problem. From project ownership and experiment trees through isolated sessions, immutable source snapshots, supervised runs, compute choices, budgeted decisions, literature, and artifacts, it connects each hypothesis to inspectable evidence. The result is a local-first control model for asking what changed, what ran, and what the record supports. It improves lineage and execution visibility, but does not guarantee valid data, meaningful metrics, identical environments, or generalizable scientific conclusions.

Learning outcomes

OpenResearch ties a hypothesis to a project, parent branch, recorded commit, run, log, and artifact, making attribution inspectable.
Private worktrees isolate agent edits while shared Git history keeps branches discoverable; they do not coordinate every research decision.
A completed or successful run is not sufficient evidence: inspect the source, command, effective configuration, metrics, and trajectory.
Autoresearch uses repair, refill, promote, and stop to turn each result into a bounded next decision.
The system strengthens lineage and execution continuity, but data validity, metric meaning, environment control, and generalization remain human judgments.

Keywords

OpenResearchreproducibilityautoresearchGitlineageagentsworktreessnapshotsevidenceartifacts

Chapters

OpenResearch: Research Reproducibility

Link ideas, commits, runs, and evidence

9 min · Link ideas, commits, runs, and evidence

9 min

OpenResearch Project Ownership

Set the repository, baseline, and run contract

8 min · Set the repository, baseline, and run contract

8 min

OpenResearch Experiment Tree

Preserve comparable hypotheses through lineage

7 min · Preserve comparable hypotheses through lineage

7 min

OpenResearch Agent Worktrees

Isolate edits while keeping branches discoverable

9 min · Isolate edits while keeping branches discoverable

9 min

OpenResearch Research Preflight

Check source, branch, session, command, and evidence

7 min · Check source, branch, session, command, and evidence

7 min

OpenResearch Run Evidence

From committed source to inspectable output

8 min · From committed source to inspectable output

8 min

OpenResearch Run Contracts

Choose local, SSH, scheduled, or ephemeral compute

9 min · Choose local, SSH, scheduled, or ephemeral compute

9 min

OpenResearch Autoresearch Loop

Repair, refill, promote, or stop on evidence

8 min · Repair, refill, promote, or stop on evidence

8 min

OpenResearch Literature and Artifacts

Keep discovery, interpretation, and presentation distinct

9 min · Keep discovery, interpretation, and presentation distinct

9 min

OpenResearch: Local Control and Remote Power

Where lineage helps and scientific guarantees stop

9 min · Where lineage helps and scientific guarantees stop

9 min

Preview every chapter · Unlock full text, audio, and video from chapter 4

Plans

Ch.1-3 FreeCh.4-10 Premium

FAQ

What does OpenResearch make reproducible?

OpenResearch makes source identity, experiment lineage, execution state, and recorded output more inspectable; it does not by itself reproduce every environment or establish scientific validity.

How does OpenResearch keep agent experiments comparable?

It uses project baselines, parent-child Git branches, fixed run commands, committed snapshots, and shared evidence expectations; these expose intended variation without making remote environments identical.

Can OpenResearch run research on remote compute?

Yes, OpenResearch can send a recorded source snapshot to local, SSH, scheduled, serverless, specialized, or managed compute through adapters, while keeping the run tied to the local experiment record.

How does OpenResearch decide what to do after a run?

After reconciling terminal runs and inspecting logs and diffs, the loop repairs unanswered nodes, refills inconclusive rounds, promotes clear winners, or stops at a goal, evidence, branch, or budget boundary.

What does OpenResearch not guarantee?

It does not guarantee representative data, meaningful metrics, identical environments, controlled randomness, secure remote infrastructure, or generalization; those require separate controls and judgment.