Awesome Loop Engineering v0.9.0
Awesome Loop Engineering v0.9.0 maps recurrence from a model's inner computation to an agent's tools and the outer operating loop that verifies, remembers, retries, escalates, and stops.
Explore 601 resources without collapsing looped models, agent execution, harnesses, workflows, and production operations into one idea.
What You Can Use
- 601 papers, official docs, tools, benchmarks, and guides linked to the original work
- 29 model-layer resources, from Universal Transformers and Huginn to Loopie, LoopCoder, and LoopWM, labeled as adjacent foundations rather than complete agent loops
- 20 operational patterns organized by build, operate, optimize, and govern use cases
- 20 schema-checked loop contracts, one for every pattern
- 8 runtime starters: 3 dependency-light executables and 5 copy/paste runtime templates
- an interactive Resource Atlas for filtering by goal, loop layer, lifecycle stage, artifact type, and evidence class
- 50-field CSV, JSONL, and Parquet exports mirrored as a Hugging Face dataset
- 8 language entry points
What Changed In v0.9.0
- Added Loopie and 21 more works across model recurrence, agent workflows, verification, security, memory, orchestration, evaluation, and operations.
- Expanded Model-Level Recurrence to 29 resources, including sparse MoE recurrence, fixed-point halting, and mechanistic evidence, with the deeper Awesome Loop Models catalog for further discovery.
- Added
loop_layerandscope_fitto every dataset row so model recurrence remains discoverable without being misclassified as operational Loop Engineering. - Added a model-to-operations map and loop-layer filter to the Resource Atlas.
- Added a cross-layer research protocol for comparing internal recurrent depth with external evidence-aware retries under matched compute and cost.
- Checked all 601 source links and refreshed every arXiv publication record so verified conference or journal publications take precedence over preprints.
- Updated the website, social preview, translations, release metadata, and Hugging Face dataset to the same counts and terminology.
Why This Matters
"Loop" is overloaded. A learned block can recur inside one inference, an agent can alternate reasoning and tools inside one task, and an operating system can rerun verified work across time. Each layer matters, but each carries different state, stopping rules, evidence, and risks.
The v0.9.0 map keeps those layers connected and comparable:
- Study model recurrence when the question is adaptive depth or latent computation.
- Study agent and harness resources when the question is reasoning, tools, context, or verification within a task.
- Use an operational pattern and Loop Contract when work must recur across events, sessions, or time.
- Let external evidence, durable state, a hard budget, and human escalation govern real-world repetition.
The goal remains bounded, reviewable, evidence-driven repetition, not unlimited autonomy.
Explore And Reuse
- Explore the Resource Atlas
- Choose an operational pattern
- Adapt a schema-checked contract
- Run a starter
- Use the Hugging Face dataset
- Contribute a source or correction
Corrections are especially valuable. If a summary is inaccurate or a better original or official source exists, use the annotation-correction form or open a pull request.