cy0307/awesome-loop-engineering

Awesome Loop Engineering Dataset

A structured dataset of 730 papers, official docs, tools, benchmarks, patterns, critiques, and implementation guides for recurring AI-agent systems.

Resource Atlas · GitHub field guide · Resource selection · Report a correction

 

Dataset Summary

Each row connects an original source to its contribution, novelty, impact, publication details, lifecycle stages, audience, evidence type, link status, and repository context when applicable.

This dataset is generated from the GitHub field guide, which grows near-daily with source-verified additions. If it is useful to you, starring the repo is the easiest way to follow new resources and helps other builders find the collection.

Current release: v0.10.0

Surface	Count
Resources	730
Research papers	284
Model-layer resources	29
Operational patterns	22
Adaptable loop contracts	22
Runtime starters	8
Dataset fields	50
Language entry points	8

At the 2026-07-25 source check, 677 public links opened successfully, 2 required access, 51 pointed to repository files, and 0 were broken or unreachable.

What A Loop Contract Is

A Loop Contract is a reviewable operating specification for one recurring agent job. It fixes what authorizes a run, which work and actions are allowed, what context and roles the run uses, which external evidence proves progress, what state survives, how much the loop may spend, and when a human takes over or the loop exits.

Recurring agents need this policy because schedules, events, queues, and goals remove live supervision. Without explicit boundaries, missing decisions become hidden defaults: the loop can select the wrong work, widen its own scope, approve itself, forget failed attempts, or retry without a stopping rule. Start with the JSON schema and adapt one of the 22 schema-checked examples.

Loop Maturity Model

The maturity model helps teams choose the smallest operating model that can perform one recurring job reliably. Levels describe how a workflow is triggered, remembered, verified, divided, and supervised; they do not score model intelligence or product quality.

Level	Added capability	Example
0 · Manual prompting	A person holds state and judges each next action.	One supervised test repair.
1 · Scripted retry	A bounded wrapper returns external failure to one agent.	Three pytest repair attempts.
2 · Scheduled loop	A schedule or event starts bounded intake and reporting.	Nightly docs-drift scan.
3 · Stateful loop	Checkpoints and receipts survive across runs.	Feedback clustering with processed IDs.
4 · Self-verifying loop	External checks or independent evaluation gate completion.	CI repair exits only when the original check passes.
5 · Multi-agent loop	Specialists use explicit handoffs and shared state.	Explorer, reproducer, and reviewer for security findings.
6 · Production-supervised loop	Telemetry, least privilege, budgets, approvals, rollback, and escalation govern production work.	Deploy verification with threshold-based pause and human rollback approval.

Move up only when the current level cannot meet a recurring continuity, evidence, or risk requirement. Durable state should precede longer unattended runs, external verification should precede more agents, and production controls should precede production impact. The full field guide explains why each level exists and the signal for advancing.

Load The Data

from datasets import load_dataset

resources = load_dataset(
    "cy0307/awesome-loop-engineering",
    "resources",
    split="train",
)

papers = resources.filter(lambda row: row["resource_type"] == "Paper")
verification = resources.filter(
    lambda row: "verification" in row["lifecycle_stages"].split(";")
)
model_recurrence = resources.filter(
    lambda row: row["loop_layer"] == "model"
    and row["scope_fit"] == "adjacent"
)

For pandas:

import pandas as pd

resources = pd.read_csv(
    "https://huggingface.co/datasets/cy0307/awesome-loop-engineering/resolve/main/data/resources.csv"
)

Use url as the durable join key. row_id and source_line are positional and can change when the README is reordered.

Intended Uses

• Find primary sources and implementation references for recurring agent systems.
• Compare works by lifecycle, audience, evidence class, source type, and publication metadata.
• Separate model, agent, harness, workflow, operations, evaluation, and cross-layer resources without treating adjacent model recurrence as a complete operational loop.
• Build literature maps, reading lists, dashboards, or retrieval indexes.
• Inspect contribution, novelty, impact, source, and evidence fields before following a claim.
• Find reusable patterns, contracts, schemas, executable examples, and copy/paste runtime templates.

Do not use signal_strength, GitHub stars, forks, or inclusion in this collection as a quality label, endorsement, or automated ranking of scientific validity.

Future Directions

The agenda organizes fifteen measurable workstreams in dependency order:

Tier	Prove next	First artifacts
Foundation	Verification, state, recovery, receipts, and security remain trustworthy under failure	Challenge sets, fault injection, replay, receipt schemas, and enforced permissions
Scale	Gains survive ablation, long horizons, matched budgets, runtime changes, and cost accounting	Factorized benchmarks, control-policy replays, contract adapters, and economic frontiers
Adoption	Operators can deploy, understand, interrupt, hand off, update, and retire a useful loop	Domain pilots, promotion gates, handoff studies, incident drills, and lifecycle controls

The complete Future Directions agenda provides a shared evaluation protocol, operational metric definitions, starter slices and completion gates for every workstream, 90-day role plans, field milestones, and a proposal template.

Dataset Structure

The primary configuration is resources, with one train split backed by the native data/resources.parquet shard. data/resources.jsonl and data/resources.csv contain the same rows for streaming, spreadsheet, and dataframe workflows.

Field group	Fields	What it describes
Identity	row_id, title, url, canonical_url, resource_type, domain	What the resource is and where it lives.
Resource analysis	annotation, key_contribution, novelty, impact	What the resource contributes to recurring agent systems.
Scope	loop_layer, scope_fit	Where recurrence lives and whether the source is direct, enabling, or adjacent to operational Loop Engineering.
Navigation	section, collection, user_goal, lifecycle_stages, audience	Where the resource fits and who it serves.
Evidence	evidence_class, evidence_tier, signal, signal_strength, source_status, audited_at	What kind of evidence or provenance is available.
Publication	authors, publication_date, publication_year, publication_venue, publisher, doi, arxiv_id, primary_category	Bibliographic data exposed by the original or official source.
Source details	source_title, source_description, metadata_source, publication_note	Where metadata came from and what caveats accompany it.
Repository context	github_repo, github_stars, github_forks, github_license, github_created_at, github_updated_at	Point-in-time adoption and maintenance context for GitHub projects.
Export trace	source_readme, source_line, source_url, date_added	How the row maps back to the full field guide.

The complete field-by-field schema is documented in data/README.md.

How To Read The Evidence

Open the linked work before relying on a summary. Evidence labels distinguish published research, official documentation, inspectable implementations, practitioner analysis, risk analysis, and discovery indexes; they describe the kind of source, not its scientific quality.

Summaries explain each work's contribution in original language rather than copying abstracts. They do not imply author endorsement. Follow the source link to inspect methods, results, limitations, licenses, and current documentation.

The latest link-check results are available as data/resource_source_audit.csv.

Limitations

• Coverage is selective rather than exhaustive, and one person currently reviews changes.
• Coverage is skewed toward English-language, publicly available material.
• A successful URL check proves that a link opened at check time, not that its claims are correct, permanent, or independently validated.
• Access-restricted rows could not be fully retrieved during the latest check and should be opened manually.
• Publication dates and authors remain blank when the primary source does not expose reliable metadata.
• Resource statistics are snapshots and will drift after the recorded audited_at timestamp.
• Linked third-party works retain their own licenses and terms; CC0-1.0 covers only original repository descriptions, metadata, templates, and documentation.

Versioning And Corrections

GitHub Releases define versioned snapshots; cite a release or commit for reproducibility. Submit corrections to summaries, contribution, novelty, impact, authorship, dates, venues, identifiers, or source URLs through the correction form.

Follow And Support

Everything here starts life in the GitHub repository — the dataset, the atlas, the patterns, and the contracts are all generated from it.

• ⭐ Star the repo to follow near-daily verified additions and help others discover the collection.
• 🍴 Fork it to adapt the 22 patterns, 22 loop contracts, and 8 runtime starters to your own stack.
• 🧭 Suggest a resource or open a PR — every submission is source-verified before it lands.
• 🤗 Liking this dataset on Hugging Face also helps other practitioners find it.

Citation

@misc{chaoyue2026awesome_loop_engineering,
  author       = {He, Chaoyue},
  title        = {Awesome Loop Engineering},
  year         = {2026},
  howpublished = {\url{https://github.com/ChaoYue0307/awesome-loop-engineering}},
  note         = {Version 0.10.0}
}