Loop Daily: April 08, 2026
The autoresearch pattern is spreading fast beyond its Karpathy-era origins. April 6 saw legal domain optimization at Harvey, a hackathon with real GPU staking, and multiple builders running 24/7 autonomous loops on everything from trading systems to personality engines. The harness itself is becoming the conversation.
#1
@gabepereyra
https://x.com/gabepereyra/status/2041167397453758863
Auto-research applied to legal agents at Harvey. The team published a summary of how they optimize agent harnesses for domain-specific tasks. This is autoresearch leaving the generic "improve my code" territory and entering specialized professional domains where the quality bar is much higher and the iteration cycles are completely different.
https://x.com/gabepereyra/status/2041167397453758863
Auto-research applied to legal agents at Harvey. The team published a summary of how they optimize agent harnesses for domain-specific tasks. This is autoresearch leaving the generic "improve my code" territory and entering specialized professional domains where the quality bar is much higher and the iteration cycles are completely different.
#2
@dgalarza
https://x.com/dgalarza/status/2041169301923643410
Used autoresearch as a Claude Code skill: 65 iterations across three rounds, fully autonomous. The loop is simple: change, commit, check metric, quality guard, keep or rollback, repeat. What makes this notable is the use of autoresearch on production code rather than toy benchmarks. The quality guard step is the key innovation that keeps the loop from optimizing toward garbage.
https://x.com/dgalarza/status/2041169301923643410
Used autoresearch as a Claude Code skill: 65 iterations across three rounds, fully autonomous. The loop is simple: change, commit, check metric, quality guard, keep or rollback, repeat. What makes this notable is the use of autoresearch on production code rather than toy benchmarks. The quality guard step is the key innovation that keeps the loop from optimizing toward garbage.
#3
@btfdandhodl
https://x.com/btfdandhodl/status/2041116624573382984
Running Qwen 27b on a single 3090 for two weeks doing Karpathy's auto-research pattern, feeding results directly into a trading system for continuous 24/7 analysis with zero API costs. Just upgraded to dual 3090s. This is the financial autoresearch use case people keep asking about: local model, zero marginal cost, always-on analysis loop feeding real trading decisions.
https://x.com/btfdandhodl/status/2041116624573382984
Running Qwen 27b on a single 3090 for two weeks doing Karpathy's auto-research pattern, feeding results directly into a trading system for continuous 24/7 analysis with zero API costs. Just upgraded to dual 3090s. This is the financial autoresearch use case people keep asking about: local model, zero marginal cost, always-on analysis loop feeding real trading decisions.
#4
@hrishioa
https://x.com/hrishioa/status/2041199021839233070
Wrote the clearest definition of what a harness actually is: a system prompt with basic tool definitions for read, write, exec, and external calls. Everything else (sandboxing, extensions, subagents, memory, guardrails) is optional. The base function of a harness is a stable agentic loop that supports installed tools without breaking. Breaking means death loops, failed edits, or broken extensions. Simple and precise.
https://x.com/hrishioa/status/2041199021839233070
Wrote the clearest definition of what a harness actually is: a system prompt with basic tool definitions for read, write, exec, and external calls. Everything else (sandboxing, extensions, subagents, memory, guardrails) is optional. The base function of a harness is a stable agentic loop that supports installed tools without breaking. Breaking means death loops, failed edits, or broken extensions. Simple and precise.
#5
@_shubhankar
https://x.com/_shubhankar/status/2041001265904418886
Running autoresearch on production browsing skills, specifically applying the concept to real production traces that run in the background. Taking autoresearch from the experiment phase to continuous production improvement is the natural evolution that most people have not made yet.
https://x.com/_shubhankar/status/2041001265904418886
Running autoresearch on production browsing skills, specifically applying the concept to real production traces that run in the background. Taking autoresearch from the experiment phase to continuous production improvement is the natural evolution that most people have not made yet.
#6
@marcopaul
https://x.com/marcopaul/status/2041031325713420311
Spent the entire Easter weekend experimenting with autoresearch from morning to evening. Calls it "truly addictive." The pattern of watching an autonomous loop slowly improve something real creates a unique kind of engagement that manual coding does not match.
https://x.com/marcopaul/status/2041031325713420311
Spent the entire Easter weekend experimenting with autoresearch from morning to evening. Calls it "truly addictive." The pattern of watching an autonomous loop slowly improve something real creates a unique kind of engagement that manual coding does not match.
#7
@marketcallsHQ
https://x.com/marketcallsHQ/status/2041187315255788028
Using vectorBT combined with autoresearch workflow to automatically improve backtesting strategies. The autoresearch loop creates the backtest with vectorBT skills, evaluates it, and iterates to find better parameters. Quantitative finance meets autonomous improvement.
https://x.com/marketcallsHQ/status/2041187315255788028
Using vectorBT combined with autoresearch workflow to automatically improve backtesting strategies. The autoresearch loop creates the backtest with vectorBT skills, evaluates it, and iterates to find better parameters. Quantitative finance meets autonomous improvement.
#8
@amenoacids
https://x.com/amenoacids/status/2041301742693105814
Improving a cross-platform personality system using pi-autoresearch. Non-coding application: the loop optimizes how a personality behaves across different platforms, not source code.
https://x.com/amenoacids/status/2041301742693105814
Improving a cross-platform personality system using pi-autoresearch. Non-coding application: the loop optimizes how a personality behaves across different platforms, not source code.
#9
@RK9409758828622
https://x.com/RK9409758828622/status/2041135757063004450
Connected autoresearch to OpenClaw and getting news notifications about how experiments are progressing. The integration means your autoresearch loop can push updates to your agent, creating a feedback channel between the autonomous experiment and your daily workflow.
https://x.com/RK9409758828622/status/2041135757063004450
Connected autoresearch to OpenClaw and getting news notifications about how experiments are progressing. The integration means your autoresearch loop can push updates to your agent, creating a feedback channel between the autonomous experiment and your daily workflow.
#10
@NaitiveAi
https://x.com/NaitiveAi/status/2041276108793135104
Autoresearch-mlx is an Apple Silicon port of Karpathy's autoresearch that runs without PyTorch. Uses program.md as the configuration file. Makes autoresearch accessible to Mac users without the CUDA dependency.
https://x.com/NaitiveAi/status/2041276108793135104
Autoresearch-mlx is an Apple Silicon port of Karpathy's autoresearch that runs without PyTorch. Uses program.md as the configuration file. Makes autoresearch accessible to Mac users without the CUDA dependency.
#11
@dave_cameron
https://x.com/dave_cameron/status/2041144018600390941
Playing with pi-autoresearch using Gemma 4. The combination of a free local model with the autoresearch pattern is exactly the zero-cost experimentation loop that makes this accessible to everyone.
https://x.com/dave_cameron/status/2041144018600390941
Playing with pi-autoresearch using Gemma 4. The combination of a free local model with the autoresearch pattern is exactly the zero-cost experimentation loop that makes this accessible to everyone.
#12
@tapansharma04
https://x.com/tapansharma04/status/2041127944932950046
Auto Kernel is an open-source framework that applies an autonomous agent loop specifically to GPU kernel optimization for arbitrary PyTorch models. The loop automatically generates and tests kernel configurations to find optimal performance. Specialized infrastructure for the ML training pipeline.
https://x.com/tapansharma04/status/2041127944932950046
Auto Kernel is an open-source framework that applies an autonomous agent loop specifically to GPU kernel optimization for arbitrary PyTorch models. The loop automatically generates and tests kernel configurations to find optimal performance. Specialized infrastructure for the ML training pipeline.
#13
@PyImageSearch
https://x.com/PyImageSearch/status/2041169228443611488
Self-correcting vision system combining Qwen model with Meta SAM 3. The iterative agent loop produces smarter segmentation by having the system critique and improve its own outputs. Computer vision meets the autoresearch pattern.
https://x.com/PyImageSearch/status/2041169228443611488
Self-correcting vision system combining Qwen model with Meta SAM 3. The iterative agent loop produces smarter segmentation by having the system critique and improve its own outputs. Computer vision meets the autoresearch pattern.
#14
@gakonst
https://x.com/gakonst/status/2041184796848844945
Autoresearch Hackathon this Thursday. Accepting applications now. Participants get stablecoins to train against real GPUs. The fact that a hackathon is now being organized around this pattern shows it has crossed from experiment to community movement.
https://x.com/gakonst/status/2041184796848844945
Autoresearch Hackathon this Thursday. Accepting applications now. Participants get stablecoins to train against real GPUs. The fact that a hackathon is now being organized around this pattern shows it has crossed from experiment to community movement.
π‘ Eco Products Radar
Eco Products Radar
pi-autoresearch β Fork/variant of autoresearch pattern, used for personality and non-code optimization
autoresearch-mlx β Apple Silicon port, no PyTorch dependency
vectorBT β Backtesting framework combined with autoresearch for quant strategies
OpenClaw β Agent framework being integrated with autoresearch loops
Auto Kernel β GPU kernel optimization via autonomous agent loop
Gemma 4 β Free local model enabling zero-cost autoresearch
pi-autoresearch β Fork/variant of autoresearch pattern, used for personality and non-code optimization
autoresearch-mlx β Apple Silicon port, no PyTorch dependency
vectorBT β Backtesting framework combined with autoresearch for quant strategies
OpenClaw β Agent framework being integrated with autoresearch loops
Auto Kernel β GPU kernel optimization via autonomous agent loop
Gemma 4 β Free local model enabling zero-cost autoresearch
Comments