Loop Daily: April 07, 2026
The autoresearch pattern is quietly leaving the lab. What started as overnight ML experiment loops is now powering GPU kernel optimization, trading strategy backtests, knowledge graph construction, binary reverse engineering, and even full web agency operations. The most interesting signal from April 5 is not any single project but a pattern shift: people are applying the keep/revert loop to things that are not code at all. Knowledge architectures, business processes, content pipelines. The loop is becoming a general-purpose improvement primitive, and the tooling around it is diversifying fast.
#1
@Akashi203
https://x.com/Akashi203/status/2040781342535790810
Published autokernel on arxiv, directly inspired by Karpathy's autoresearch. Applied the keep/revert agent loop to GPU kernel optimization and the numbers are real: 5.29x over PyTorch eager on RMSNorm, 2.82x on Softmax, beats torch.compile by 3.44x. Grabbed the number one spot on the vectorsum_v2 B200 leaderboard. A single-prompt Triton FP4 matmul beats CUTLASS by up to 2.15x. The system churns through roughly 40 experiments per hour overnight with zero human intervention.
https://x.com/Akashi203/status/2040781342535790810
Published autokernel on arxiv, directly inspired by Karpathy's autoresearch. Applied the keep/revert agent loop to GPU kernel optimization and the numbers are real: 5.29x over PyTorch eager on RMSNorm, 2.82x on Softmax, beats torch.compile by 3.44x. Grabbed the number one spot on the vectorsum_v2 B200 leaderboard. A single-prompt Triton FP4 matmul beats CUTLASS by up to 2.15x. The system churns through roughly 40 experiments per hour overnight with zero human intervention.
#2
@0xSero
https://x.com/0xSero/status/2040731819125981515
Argues that autoresearch hype has died down but the underlying paradigm is real and durable. The claim is that millions of things worth money, attention, and recognition can be autoresearched today. This is the contrarian bet: the pattern is underhyped relative to its actual utility, not overhyped. Most people moved on to the next shiny thing while the real practitioners are quietly running loops every night.
https://x.com/0xSero/status/2040731819125981515
Argues that autoresearch hype has died down but the underlying paradigm is real and durable. The claim is that millions of things worth money, attention, and recognition can be autoresearched today. This is the contrarian bet: the pattern is underhyped relative to its actual utility, not overhyped. Most people moved on to the next shiny thing while the real practitioners are quietly running loops every night.
#3
@marketcallsHQ
https://x.com/marketcallsHQ/status/2040727182453231671
Built a self-improving AI backtesting workflow using Claude Code, OpenAlgo, DuckDB, and VectorBT. This is the autoresearch loop applied to quantitative trading with an EMA crossover strategy as the demo case. The system iterates on strategy parameters autonomously. Trading is one of those domains where the fitness function is obvious (profit) but the search space is enormous, making it a natural fit for agent loops.
https://x.com/marketcallsHQ/status/2040727182453231671
Built a self-improving AI backtesting workflow using Claude Code, OpenAlgo, DuckDB, and VectorBT. This is the autoresearch loop applied to quantitative trading with an EMA crossover strategy as the demo case. The system iterates on strategy parameters autonomously. Trading is one of those domains where the fitness function is obvious (profit) but the search space is enormous, making it a natural fit for agent loops.
#4
@ihtesham2005
https://x.com/ihtesham2005/status/2040776552515088637
Released AutoAgent, a meta-agent that builds, tests, scores, and improves AI agents in a loop. Everything is controlled by a single program.md file. Harbor-compatible benchmarks, MIT license. The interesting architectural choice is treating agent-building itself as the optimization target. Instead of the loop improving code or a model, it improves the agent that will then do the work.
https://x.com/ihtesham2005/status/2040776552515088637
Released AutoAgent, a meta-agent that builds, tests, scores, and improves AI agents in a loop. Everything is controlled by a single program.md file. Harbor-compatible benchmarks, MIT license. The interesting architectural choice is treating agent-building itself as the optimization target. Instead of the loop improving code or a model, it improves the agent that will then do the work.
#5
@RoscoeSitePro
https://x.com/RoscoeSitePro/status/2040693678163767752
Running an AI agent as a full web agency. The workspace is structured as a self-maintaining knowledge base covering prospects, clients, analytics, competitive intel, and logs. The agent runs nightly health checks, finds the weakest revenue link, runs experiments, logs results. This is the autoresearch pattern applied to business operations rather than technical research. The loop target is revenue, not accuracy.
https://x.com/RoscoeSitePro/status/2040693678163767752
Running an AI agent as a full web agency. The workspace is structured as a self-maintaining knowledge base covering prospects, clients, analytics, competitive intel, and logs. The agent runs nightly health checks, finds the weakest revenue link, runs experiments, logs results. This is the autoresearch pattern applied to business operations rather than technical research. The loop target is revenue, not accuracy.
#6
@Grisokay
https://x.com/Grisokay/status/2040601062038700256
Built a "Subconscious agent" that runs as a continuous LLM process hunting for useful problems in the background. The flow is IDEA to CHALLENGE to DEFEND to REVISE to REJECT/ACCEPT/SHELVE, with a maximum of 3 challenge rounds per idea. Runs on local Qwen3.5 9B plus ChatGPT 5.4 mini via Hermes. Think of it as autoresearch applied to idea generation itself rather than code improvement. The agent is not optimizing a metric; it is generating and stress-testing hypotheses.
https://x.com/Grisokay/status/2040601062038700256
Built a "Subconscious agent" that runs as a continuous LLM process hunting for useful problems in the background. The flow is IDEA to CHALLENGE to DEFEND to REVISE to REJECT/ACCEPT/SHELVE, with a maximum of 3 challenge rounds per idea. Runs on local Qwen3.5 9B plus ChatGPT 5.4 mini via Hermes. Think of it as autoresearch applied to idea generation itself rather than code improvement. The agent is not optimizing a metric; it is generating and stress-testing hypotheses.
#7
@morganlinton
https://x.com/morganlinton/status/2040810925004104079
Open-sourced an autoresearch project in Rust. Playing with the concept on a real codebase and sharing from MVP onward. Notable for being one of the few Rust implementations of the pattern, inviting community collaboration from day one rather than polishing in private first.
https://x.com/morganlinton/status/2040810925004104079
Open-sourced an autoresearch project in Rust. Playing with the concept on a real codebase and sharing from MVP onward. Notable for being one of the few Rust implementations of the pattern, inviting community collaboration from day one rather than polishing in private first.
#8
@kobi_gg
https://x.com/kobi_gg/status/2040788798775177387
Applied the autoresearch pattern to a knowledge system rather than model training. No GPU needed, just a laptop plus Neo4j plus Claude API. Total cost: 17 dollars. The propose-change, evaluate, keep/revert cycle works on knowledge architecture the same way it works on code. This is one of the clearest demonstrations that the loop pattern has nothing inherently to do with ML. It is a general optimization primitive.
https://x.com/kobi_gg/status/2040788798775177387
Applied the autoresearch pattern to a knowledge system rather than model training. No GPU needed, just a laptop plus Neo4j plus Claude API. Total cost: 17 dollars. The propose-change, evaluate, keep/revert cycle works on knowledge architecture the same way it works on code. This is one of the clearest demonstrations that the loop pattern has nothing inherently to do with ML. It is a general optimization primitive.
#9
@mamagnus00
https://x.com/mamagnus00/status/2040697119972196407
Called out Codex as horrible for autoresearch because it stops after a few turns if you tell it to run in a loop overnight. The workaround: explain one step of your loop and queue hundreds of them individually. 249 likes and 33K impressions suggest this is a pain point many people share. The infrastructure gap between what people want to do with agent loops and what current tools support is real and persistent.
https://x.com/mamagnus00/status/2040697119972196407
Called out Codex as horrible for autoresearch because it stops after a few turns if you tell it to run in a loop overnight. The workaround: explain one step of your loop and queue hundreds of them individually. 249 likes and 33K impressions suggest this is a pain point many people share. The infrastructure gap between what people want to do with agent loops and what current tools support is real and persistent.
#10
@EmergentMind
https://x.com/EmergentMind/status/2040797574269997416
Published the Omni-SimpleMem paper showing autoresearch-guided discovery of lifelong multimodal agent memory. The AI autonomously ran a multistage research loop, diagnosed runtime errors mid-experiment, swapped embedding models on the fly, and achieved a 411 percent F1 score improvement. This is autoresearch doing actual research, not just parameter tuning. The agent made architectural decisions during the loop.
https://x.com/EmergentMind/status/2040797574269997416
Published the Omni-SimpleMem paper showing autoresearch-guided discovery of lifelong multimodal agent memory. The AI autonomously ran a multistage research loop, diagnosed runtime errors mid-experiment, swapped embedding models on the fly, and achieved a 411 percent F1 score improvement. This is autoresearch doing actual research, not just parameter tuning. The agent made architectural decisions during the loop.
#11
@ramirosalas
https://x.com/ramirosalas/status/2040678105573146798
Claims to have been doing this long before it was called autoresearch, evolving algorithms with GEPA along the Pareto frontier. Discovering new alpha every night. This reinforces the pattern convergence thesis: multiple independent practitioners arrived at the same core loop from completely different starting points. The name is new, the practice is not.
https://x.com/ramirosalas/status/2040678105573146798
Claims to have been doing this long before it was called autoresearch, evolving algorithms with GEPA along the Pareto frontier. Discovering new alpha every night. This reinforces the pattern convergence thesis: multiple independent practitioners arrived at the same core loop from completely different starting points. The name is new, the practice is not.
#12
@_chinmaymk
https://x.com/_chinmaymk/status/2040627180808609890
Built a custom tool for autoresearch and pushed F1 score up by 10 points. The insight: once you have autoresearch you also need a code reviewer and planner, so one integrated tool handles all three roles. The loop alone is not enough. You need the scaffolding around it to prevent quality degradation over many iterations.
https://x.com/_chinmaymk/status/2040627180808609890
Built a custom tool for autoresearch and pushed F1 score up by 10 points. The insight: once you have autoresearch you also need a code reviewer and planner, so one integrated tool handles all three roles. The loop alone is not enough. You need the scaffolding around it to prevent quality degradation over many iterations.
#13
@analukach
https://x.com/analukach/status/2040927884123279670
Combined autoresearch with a knowledge database in Obsidian using a "tree-research" variant. Parallel research branches produce real insights. The claim is it works like a human brain, remembering and cross-referencing findings across branches. This is the autoresearch loop meeting personal knowledge management, two trends colliding in a way that makes each stronger.
https://x.com/analukach/status/2040927884123279670
Combined autoresearch with a knowledge database in Obsidian using a "tree-research" variant. Parallel research branches produce real insights. The claim is it works like a human brain, remembering and cross-referencing findings across branches. This is the autoresearch loop meeting personal knowledge management, two trends colliding in a way that makes each stronger.
#14
@androolloyd
https://x.com/androolloyd/status/2040766790519726101
Running a Claude-and-Claude automated flow for binary reverse engineering. Full dependency graph reconstruction, byte-by-byte scanning, trace analysis. Says he is not really involved in the process at all, just drinks coffee and runs his company. This is the most literal version of the overnight agent loop applied to a deeply technical domain that normally requires specialized human expertise.
https://x.com/androolloyd/status/2040766790519726101
Running a Claude-and-Claude automated flow for binary reverse engineering. Full dependency graph reconstruction, byte-by-byte scanning, trace analysis. Says he is not really involved in the process at all, just drinks coffee and runs his company. This is the most literal version of the overnight agent loop applied to a deeply technical domain that normally requires specialized human expertise.
#15
@010Zaj
https://x.com/010Zaj/status/2040935530305912871
Running autoresearch with OpenClaw, where native video generation opens up data visualizations, experiment result animations, and auto-generated report videos. This extends the loop output beyond text and code into multimedia, which matters for communicating results to non-technical stakeholders.
https://x.com/010Zaj/status/2040935530305912871
Running autoresearch with OpenClaw, where native video generation opens up data visualizations, experiment result animations, and auto-generated report videos. This extends the loop output beyond text and code into multimedia, which matters for communicating results to non-technical stakeholders.
#16
@krishna18421
https://x.com/krishna18421/status/2040813984312619118
Mapped the entire Agentic AI extension stack: Skills for WHAT to know, MCP for HOW to connect, Subagents for WHO does work, Hooks for WHEN to automate, CLAUDE.md for WHERE to ground, Plugins for HOW to ship. 148 likes and 3.8K impressions. This is less a project and more a mental model for understanding where agent loops fit in the broader infrastructure picture.
https://x.com/krishna18421/status/2040813984312619118
Mapped the entire Agentic AI extension stack: Skills for WHAT to know, MCP for HOW to connect, Subagents for WHO does work, Hooks for WHEN to automate, CLAUDE.md for WHERE to ground, Plugins for HOW to ship. 148 likes and 3.8K impressions. This is less a project and more a mental model for understanding where agent loops fit in the broader infrastructure picture.
#17
@Dallenpyrah
https://x.com/Dallenpyrah/status/2040623813055086774
Uses autoresearch on loop to reduce lines of code in a codebase while maintaining all features. Does it at least once a week to tame the slop. This is the most pragmatic application of the pattern: not research, not optimization, just keeping your codebase clean by letting an agent refactor overnight. Maintenance as a loop target.
https://x.com/Dallenpyrah/status/2040623813055086774
Uses autoresearch on loop to reduce lines of code in a codebase while maintaining all features. Does it at least once a week to tame the slop. This is the most pragmatic application of the pattern: not research, not optimization, just keeping your codebase clean by letting an agent refactor overnight. Maintenance as a loop target.
#18
@paul_cal
https://x.com/paul_cal/status/2040716814976532530
Uses a specific number of diff approaches tracked in git plus a research log, or an explicit local time deadline, to bound autoresearch runs. Queued messages ask the agent to keep going until the condition is met. This is practical loop hygiene: without clear stopping conditions, agent loops either quit too early or run forever burning tokens. The craft is in the constraints.
https://x.com/paul_cal/status/2040716814976532530
Uses a specific number of diff approaches tracked in git plus a research log, or an explicit local time deadline, to bound autoresearch runs. Queued messages ask the agent to keep going until the condition is met. This is practical loop hygiene: without clear stopping conditions, agent loops either quit too early or run forever burning tokens. The craft is in the constraints.
π‘ Eco Products Radar
Eco Products Radar
OpenClaw: Open-source agent platform mentioned in multiple loop contexts. Used for self-improving agent setups with LEARNINGS.md patterns, and now supporting native video generation for autoresearch output visualization.
Hermes: NousResearch's open-source agent framework. Used as the routing layer for the Subconscious agent running local Qwen plus remote ChatGPT. Expanding to support general research tasks beyond code optimization.
Obsidian: Personal knowledge management tool now being combined with autoresearch through tree-research variants. Parallel research branches with cross-referencing create a persistent knowledge layer for loop outputs.
Claude Code: Showing up as the core engine in trading backtests, binary reverse engineering, and knowledge graph construction. The most frequently chosen agent runtime for non-ML autoresearch applications.
Codex: Called out for poor support of long-running loops. Stops after a few turns, forcing users to queue hundreds of individual steps as a workaround. The gap between what builders want and what Codex supports is a recurring frustration.
Neo4j: Graph database used as the knowledge backend for a 17-dollar autoresearch run on knowledge architecture. Demonstrates that the loop pattern works without GPU infrastructure when the optimization target is structure rather than compute.
Qwen: Qwen3.5 9B running locally as part of the Subconscious agent's idea generation pipeline. Paired with remote models via Hermes for a hybrid local/cloud setup.
Triton/CUDA: The target output format for autokernel's GPU optimization loop. Triton FP4 matmul from a single prompt beating CUTLASS shows what the keep/revert pattern can achieve on low-level kernel code.
OpenClaw: Open-source agent platform mentioned in multiple loop contexts. Used for self-improving agent setups with LEARNINGS.md patterns, and now supporting native video generation for autoresearch output visualization.
Hermes: NousResearch's open-source agent framework. Used as the routing layer for the Subconscious agent running local Qwen plus remote ChatGPT. Expanding to support general research tasks beyond code optimization.
Obsidian: Personal knowledge management tool now being combined with autoresearch through tree-research variants. Parallel research branches with cross-referencing create a persistent knowledge layer for loop outputs.
Claude Code: Showing up as the core engine in trading backtests, binary reverse engineering, and knowledge graph construction. The most frequently chosen agent runtime for non-ML autoresearch applications.
Codex: Called out for poor support of long-running loops. Stops after a few turns, forcing users to queue hundreds of individual steps as a workaround. The gap between what builders want and what Codex supports is a recurring frustration.
Neo4j: Graph database used as the knowledge backend for a 17-dollar autoresearch run on knowledge architecture. Demonstrates that the loop pattern works without GPU infrastructure when the optimization target is structure rather than compute.
Qwen: Qwen3.5 9B running locally as part of the Subconscious agent's idea generation pipeline. Paired with remote models via Hermes for a hybrid local/cloud setup.
Triton/CUDA: The target output format for autokernel's GPU optimization loop. Triton FP4 matmul from a single prompt beating CUTLASS shows what the keep/revert pattern can achieve on low-level kernel code.
Comments