Agentic Research Directory

Framework

OpenLabs: bounded research-runtime reference architecture

A public reference architecture describes recoverable, resource-bounded agent research attempts across separate domain laboratories and private research state.

Workflow level n/a Source: Repository-documented Outcome: Not empirically evaluated here

Research applicability

A reference architecture for bounded research orchestration, evidence tracking and recovery.

Getting started

Read the linked primary sources and the stated adoption requirements. Reuse suggestions are editorial and untested; no setup recipe or trial was executed.

Maintenance

Sources last checked on 2026-10-19. Future availability, maintenance, provider terms and compatibility were not verified.

License

The README states Apache 2.0 for original code and separate vendor terms; licence compatibility was not audited.

Workflow & human oversight

No workflow level assigned. The architecture documents seeded, bounded research attempts, agent-chosen scientific steps, leases, resource controls, evidence gates and recovery. Its independent reviewers are fresh agent/provider sessions; human validation of scientific conclusions is not established. Hashes and a needs-human stop do not establish L3/L4 oversight.

Workflow level is not an evidence score; higher is not better evidence.

Approach

A documented control plane tracks leases, attempt identity, resources, evidence hashes and transactional promotion while agents choose scientific routes and propose successors. Fresh-session reviews and terminal conditions are described. Public code is separate from private campaign and artefact state.

Limitations

README and runtime/constraint documents only. Code, hooks, private campaigns, artefacts, databases, claimed papers/results and services were not inspected or run. The source describes broad runtime permissions; an attempt working directory is a convention, not kernel-enforced write isolation, and resource limits are not security containment. Public code alone does not reproduce private campaign state. Hash/replay checks and fresh agent reviewers do not prove scientific truth or human acceptance. Internal readiness is not submission or authorship approval. No source permission or runtime setting was applied. Moving HEAD documents are unpinned and non-atomic.

Reproducibility

Source text was checked only within this scope: Full README and selected runtime/minimal-constraint documents. No software, dataset, experiment or result was independently reproduced.

Adoption requirements

Untested architecture reference: establish enforceable access boundaries, required private-state availability and explicit human conclusion acceptance before a trial; do not inherit source runtime permissions.

Comments

No comments yet.

Add a comment

Comments are held for moderation before appearing. Critical or corrective comments are welcome — especially evidence that an entry overstates what has been verified.

Submitting stores your comment text, the optional name and email, and a pseudonymised connection identifier used only for spam limiting (the raw IP address is not stored).