Loop Daily: 2026-04-30
If yesterday's Twitter is right, autoresearch quietly graduated from "Karpathy posted a loop a few weeks ago" to "Mac Ryzens running 10,000 iterations overnight, npm-installable harnesses, and at least one quant fund using it on production trading code." The shape of the keepers: the framework matured (pi-autoresearch hit npm, gnhf hit 1K stars), and the use cases drifted hard outside coding β investment research, biology, drone optimization, even social-media content engines. The boring observation that nobody is making yet: people are no longer pitching agentic loops, they're shipping concrete results from them.
#1
@MatthewBerman
https://x.com/MatthewBerman/status/2049195091244589252
Nightshift: an overnight autonomous ML research harness for MLX on Apple Silicon. Think AutoResearch but for laptops on a power outlet. The point isn't the tool, it's the message: anyone with a Mac can now run multi-hour ML experiments against their own ideas without renting a GPU. Karpathy's autoresearch primitive has gone consumer.
https://x.com/MatthewBerman/status/2049195091244589252
Nightshift: an overnight autonomous ML research harness for MLX on Apple Silicon. Think AutoResearch but for laptops on a power outlet. The point isn't the tool, it's the message: anyone with a Mac can now run multi-hour ML experiments against their own ideas without renting a GPU. Karpathy's autoresearch primitive has gone consumer.
#2
@davebcn87
https://x.com/davebcn87/status/2049141151484047699
pi-autoresearch can now run indefinitely. The trick: pi handles its own context compaction, then uses persisted files to pick up where it left off and keep testing new hypotheses. This is the missing piece that turns "long-running agent" from a vibe into a reliable engineering primitive β context compaction owned by the agent, not the user.
https://x.com/davebcn87/status/2049141151484047699
pi-autoresearch can now run indefinitely. The trick: pi handles its own context compaction, then uses persisted files to pick up where it left off and keep testing new hypotheses. This is the missing piece that turns "long-running agent" from a vibe into a reliable engineering primitive β context compaction owned by the agent, not the user.
#3
@davebcn87
https://x.com/davebcn87/status/2049064153730490469
pi-autoresearch shipped on npm. One install, one CLI command. The framework that everyone was hand-rolling two months ago is now apt-get-style. 624 likes / 32K impressions in a few hours β distribution mode for autoresearch is officially "package manager."
https://x.com/davebcn87/status/2049064153730490469
pi-autoresearch shipped on npm. One install, one CLI command. The framework that everyone was hand-rolling two months ago is now apt-get-style. 624 likes / 32K impressions in a few hours β distribution mode for autoresearch is officially "package manager."
#4
@0xSero
https://x.com/0xSero/status/2049048462956642620
Used autoresearch overnight to take deepseek-v4-flash on sglang from 40 tok/s to 100+ tok/s on Blackwell 6000s β 2.5Γ speedup. This is the single most concrete production result in the dataset: the loop ran while a human slept and produced measurable inference-tuning gains that would have taken an engineer days.
https://x.com/0xSero/status/2049048462956642620
Used autoresearch overnight to take deepseek-v4-flash on sglang from 40 tok/s to 100+ tok/s on Blackwell 6000s β 2.5Γ speedup. This is the single most concrete production result in the dataset: the loop ran while a human slept and produced measurable inference-tuning gains that would have taken an engineer days.
#5
@francescofaenzi
https://x.com/francescofaenzi/status/2048998654846066840
The clearest non-coding case of the day: Karpathy's autoresearch repo, originally meant for LLM architecture optimization, retargeted at quantitative finance β automated backtesting, parameter tuning, model evolution β running on Google Colab + Gemini 3.1 Pro inside a $20/mo Google One subscription. Zero additional infra cost. If this works, it commodifies systematic-trading R&D.
https://x.com/francescofaenzi/status/2048998654846066840
The clearest non-coding case of the day: Karpathy's autoresearch repo, originally meant for LLM architecture optimization, retargeted at quantitative finance β automated backtesting, parameter tuning, model evolution β running on Google Colab + Gemini 3.1 Pro inside a $20/mo Google One subscription. Zero additional infra cost. If this works, it commodifies systematic-trading R&D.
#6
@kunchenguid
https://x.com/kunchenguid/status/2048978455107383456
gnhf β author's open-source tool to run Karpathy's autoresearch on any project β hit 1,000 GitHub stars in a month. Stars accumulated organically, not bought. The fact that an autoresearch wrapper crossed 1K stars is itself the signal: this is a shipping ecosystem, not a rumor.
https://x.com/kunchenguid/status/2048978455107383456
gnhf β author's open-source tool to run Karpathy's autoresearch on any project β hit 1,000 GitHub stars in a month. Stars accumulated organically, not bought. The fact that an autoresearch wrapper crossed 1K stars is itself the signal: this is a shipping ecosystem, not a rumor.
#7
@alokbishoyi97
https://x.com/alokbishoyi97/status/2049177888902234174
Open-sourced an autoresearch orchestrator that runs on top of Claude Code or Codex with two CLI commands. Distribution play: ride the existing harnesses rather than build a new one. The path-of-least-resistance answer to "how do I start" β install on top of what you already have.
https://x.com/alokbishoyi97/status/2049177888902234174
Open-sourced an autoresearch orchestrator that runs on top of Claude Code or Codex with two CLI commands. Distribution play: ride the existing harnesses rather than build a new one. The path-of-least-resistance answer to "how do I start" β install on top of what you already have.
#8
@lftherios
https://x.com/lftherios/status/2049172075181494762
Shipped a pi-autoresearch extension and CLI that makes the runtime collaborative β multi-user, multi-agent. Single-user autoresearch is interesting; multi-agent autoresearch is where teams will actually adopt it. The collaborative layer is a real gap right now.
https://x.com/lftherios/status/2049172075181494762
Shipped a pi-autoresearch extension and CLI that makes the runtime collaborative β multi-user, multi-agent. Single-user autoresearch is interesting; multi-agent autoresearch is where teams will actually adopt it. The collaborative layer is a real gap right now.
#9
@bigmarvin
https://x.com/bigmarvin/status/2049180203063140617
Practical autoresearch tip: give your agent direct arxiv access via the arxiv-radar MCP server. Similarity search, always up to date, full papers in markdown. Setup is one line in Claude Code. Most autoresearch failures are read-the-literature failures β this fixes that loop.
https://x.com/bigmarvin/status/2049180203063140617
Practical autoresearch tip: give your agent direct arxiv access via the arxiv-radar MCP server. Similarity search, always up to date, full papers in markdown. Setup is one line in Claude Code. Most autoresearch failures are read-the-literature failures β this fixes that loop.
#10
@mjamei
https://x.com/mjamei/status/2049134291683094868
"Plug it into your autoresearch loop. Wake up to 100s of validated iterations." Plus a CI/CD deploy gate to block regressions, plus side-by-side run comparison to A/B model and harness designs. The thread reads like the operating manual for autoresearch in production engineering β overnight iteration, regression gating, A/B compare, the whole stack.
https://x.com/mjamei/status/2049134291683094868
"Plug it into your autoresearch loop. Wake up to 100s of validated iterations." Plus a CI/CD deploy gate to block regressions, plus side-by-side run comparison to A/B model and harness designs. The thread reads like the operating manual for autoresearch in production engineering β overnight iteration, regression gating, A/B compare, the whole stack.
#11
@arv_puthucode
https://x.com/arv_puthucode/status/2049185568383516680
Day 1 of a public 30-day build challenge: shipped v0.5 autoresearch mode for ScoutFox. Founders publishing daily build logs that include autoresearch progress is itself a sign β autoresearch went from research toy to founder-vocabulary in a few weeks.
https://x.com/arv_puthucode/status/2049185568383516680
Day 1 of a public 30-day build challenge: shipped v0.5 autoresearch mode for ScoutFox. Founders publishing daily build logs that include autoresearch progress is itself a sign β autoresearch went from research toy to founder-vocabulary in a few weeks.
#12
@GopiVikranth
https://x.com/GopiVikranth/status/2049143600207081647
DataClaw β an OpenClaw-based data science harness combining Hermes for the agent layer, GBrain for memory, and AutoResearch for bounded improvement loops. Tested on Kaggle datasets (warehouse retail EDA, Starbucks survey). The compose-the-pieces pattern is the news: nobody is building a single magic bullet, they're stitching ecosystem primitives.
https://x.com/GopiVikranth/status/2049143600207081647
DataClaw β an OpenClaw-based data science harness combining Hermes for the agent layer, GBrain for memory, and AutoResearch for bounded improvement loops. Tested on Kaggle datasets (warehouse retail EDA, Starbucks survey). The compose-the-pieces pattern is the news: nobody is building a single magic bullet, they're stitching ecosystem primitives.
#13
@BazarovNic39426
https://x.com/BazarovNic39426/status/2049178969287651674
Used Codex on Karpathy's autoresearch pattern β no extra tools, just told it to work as long as possible. First runs were 15-20 minutes, then 2-3 hours, now 8+ hours, almost 12 hours overnight in a single chat. The model is staying coherent across 12 hours. That's the breakthrough nobody is screenshotting.
https://x.com/BazarovNic39426/status/2049178969287651674
Used Codex on Karpathy's autoresearch pattern β no extra tools, just told it to work as long as possible. First runs were 15-20 minutes, then 2-3 hours, now 8+ hours, almost 12 hours overnight in a single chat. The model is staying coherent across 12 hours. That's the breakthrough nobody is screenshotting.
#14
@UsernameAndStuf
https://x.com/UsernameAndStuf/status/2048967226364235805
Running 10,000 recursive loops in 3.4 hours on a four-core AMD Ryzen 5 β a genetic evolution system feeding into Karpathy's autoresearch loop, using micrograd to iterate. A modified Kronos predictor LLM validates and improves the strategies. Consumer-grade hardware running thousands of evolutionary research iterations on a weekday β the cost floor is gone.
https://x.com/UsernameAndStuf/status/2048967226364235805
Running 10,000 recursive loops in 3.4 hours on a four-core AMD Ryzen 5 β a genetic evolution system feeding into Karpathy's autoresearch loop, using micrograd to iterate. A modified Kronos predictor LLM validates and improves the strategies. Consumer-grade hardware running thousands of evolutionary research iterations on a weekday β the cost floor is gone.
#15
@derekmeegan
https://x.com/derekmeegan/status/2049218109807198331
Released the /browser-trace skill: dumps network requests, DOM content, screenshots, and CDP logs into a searchable filesystem. Explicitly pitched as great for autoresearch loops and "monitoring the situation." 2,465 likes / 212K impressions β the biggest loop-related post of the day. Browser observability for agents is the new hot primitive.
https://x.com/derekmeegan/status/2049218109807198331
Released the /browser-trace skill: dumps network requests, DOM content, screenshots, and CDP logs into a searchable filesystem. Explicitly pitched as great for autoresearch loops and "monitoring the situation." 2,465 likes / 212K impressions β the biggest loop-related post of the day. Browser observability for agents is the new hot primitive.
#16
@ta_eis_eauton
https://x.com/ta_eis_eauton/status/2049155214284931530
Boiled autoresearch down to three lines: define a workspace for codex (the editable surface), define an evaluation metric, loop codex to optimize it. The whole pattern fits in a tweet. Anyone who internalizes this can build their own autoresearch loop today β no framework needed.
https://x.com/ta_eis_eauton/status/2049155214284931530
Boiled autoresearch down to three lines: define a workspace for codex (the editable surface), define an evaluation metric, loop codex to optimize it. The whole pattern fits in a tweet. Anyone who internalizes this can build their own autoresearch loop today β no framework needed.
#17
@HenryL_AI
https://x.com/HenryL_AI/status/2049272714473648315
A sharp negative result: autoresearch fails on tasks where the model has no inherent taste. Tested on Texas Hold'em β model couldn't intuit hand probabilities, so self-critique went nowhere. The fix was letting the model write an external equity solver. Pattern: when the model can't grade its own outputs, give it a tool that can.
https://x.com/HenryL_AI/status/2049272714473648315
A sharp negative result: autoresearch fails on tasks where the model has no inherent taste. Tested on Texas Hold'em β model couldn't intuit hand probabilities, so self-critique went nowhere. The fix was letting the model write an external equity solver. Pattern: when the model can't grade its own outputs, give it a tool that can.
#18
@HenryL_AI
https://x.com/HenryL_AI/status/2049272712653426995
Companion thread: LLMs aren't actually weak at evaluation in many domains β they're used as judges, critics, and self-correctors all the time. Self-evolving agents work specifically because the model evaluates its own trajectories and writes code to patch its own scaffolding. Pushes back on the "LLMs can't evaluate" critique with hard production examples.
https://x.com/HenryL_AI/status/2049272712653426995
Companion thread: LLMs aren't actually weak at evaluation in many domains β they're used as judges, critics, and self-correctors all the time. Self-evolving agents work specifically because the model evaluates its own trajectories and writes code to patch its own scaffolding. Pushes back on the "LLMs can't evaluate" critique with hard production examples.
#19
@marktenenholtz
https://x.com/marktenenholtz/status/2049147031541911663
"Autoresearch is great for training models but LOOK AT YOUR DATA. PLEASE." 20 likes / 2K impressions of pure conviction. The killer point: the smartest autoresearch loop in the world will compound errors if the underlying data is broken. Older data-science discipline applied to the new toy.
https://x.com/marktenenholtz/status/2049147031541911663
"Autoresearch is great for training models but LOOK AT YOUR DATA. PLEASE." 20 likes / 2K impressions of pure conviction. The killer point: the smartest autoresearch loop in the world will compound errors if the underlying data is broken. Older data-science discipline applied to the new toy.
#20
@derekmeegan
https://x.com/derekmeegan/status/2049262243658015091
Reply: "this is really good for creating data for auto research loops." Confirms a pattern β the bottleneck for autoresearch isn't the loop itself, it's clean training data. Tools that produce structured data (like /browser-trace) are now upstream-feeders to autoresearch pipelines.
https://x.com/derekmeegan/status/2049262243658015091
Reply: "this is really good for creating data for auto research loops." Confirms a pattern β the bottleneck for autoresearch isn't the loop itself, it's clean training data. Tools that produce structured data (like /browser-trace) are now upstream-feeders to autoresearch pipelines.
#21
@0xMovez
https://x.com/0xMovez/status/2049175936562614654
getRoman β a Slack-orchestrator agent for Polymarket trading bots. Self-improvement layer: scrape weather bot trades β analyze with Opus 4.7 β compare with 9 weather APIs β push new instructions to TG bots β self-improve from new trades. Result: weekly bot ROI went from 53% β 110%. Concrete production loop with measured deltas.
https://x.com/0xMovez/status/2049175936562614654
getRoman β a Slack-orchestrator agent for Polymarket trading bots. Self-improvement layer: scrape weather bot trades β analyze with Opus 4.7 β compare with 9 weather APIs β push new instructions to TG bots β self-improve from new trades. Result: weekly bot ROI went from 53% β 110%. Concrete production loop with measured deltas.
#22
@rblalock (linked thread by @rblalock)
https://x.com/rblalock/status/2049156828853137637
Long thread arguing the orchestration-theater era of agents is ending β Linear chains with LLM calls are being replaced by real autonomous loops as models get better at planning, tool use, and recovery. Anthropic's old "you may not need agentic systems at all" line is now the line in the sand. Useful framework for the year-on-year shift.
https://x.com/rblalock/status/2049156828853137637
Long thread arguing the orchestration-theater era of agents is ending β Linear chains with LLM calls are being replaced by real autonomous loops as models get better at planning, tool use, and recovery. Anthropic's old "you may not need agentic systems at all" line is now the line in the sand. Useful framework for the year-on-year shift.
#23
@stuffyokodraws
https://x.com/stuffyokodraws/status/2048963246485540866
Best mental model of the day: the Tao Te Ching maps to agent design. "Dao produces One, One produces Two, Two produces Three, Three produces all things." Foundation model = One; model writes own tools = Two; agent loop = Three; AGI = all things. Takeaway: don't overcomplicate the harness.
https://x.com/stuffyokodraws/status/2048963246485540866
Best mental model of the day: the Tao Te Ching maps to agent design. "Dao produces One, One produces Two, Two produces Three, Three produces all things." Foundation model = One; model writes own tools = Two; agent loop = Three; AGI = all things. Takeaway: don't overcomplicate the harness.
#24
@31Carlton7
https://x.com/31Carlton7/status/2049055506636189835
Tabs are quietly dying as the unit of software. In an agent-native app the loop is always running β observing, detecting, prioritizing, executing, narrating. Cursor's chat panel eating the editor, Claude Code as a process not a page, v0's generative canvas. Tabs as a metaphor are over; the new unit is the task with a hot loop in the background.
https://x.com/31Carlton7/status/2049055506636189835
Tabs are quietly dying as the unit of software. In an agent-native app the loop is always running β observing, detecting, prioritizing, executing, narrating. Cursor's chat panel eating the editor, Claude Code as a process not a page, v0's generative canvas. Tabs as a metaphor are over; the new unit is the task with a hot loop in the background.
#25
@Provenancetags
https://x.com/Provenancetags/status/2049018608517316953
Three layers most teams collapse: MCP (how tools connect), abstraction (which tools the harness exposes), skills (how well the model uses them). The mistake is exposing every vendor's MCP tool to the LLM β a CRM agent ends up with 36 redundant tools. Wrap them: model sees one capability, vendor routing happens below the line. Block demoed this collapsing 200+ Square endpoints to 3.
https://x.com/Provenancetags/status/2049018608517316953
Three layers most teams collapse: MCP (how tools connect), abstraction (which tools the harness exposes), skills (how well the model uses them). The mistake is exposing every vendor's MCP tool to the LLM β a CRM agent ends up with 36 redundant tools. Wrap them: model sees one capability, vendor routing happens below the line. Block demoed this collapsing 200+ Square endpoints to 3.
#26
@Everlier
https://x.com/Everlier/status/2049066541568729453
Jitera launched on Product Hunt after 5 different incarnations. Now: any agent can switch between Jitera's custom agentic loop, Codex, or Claude Code; replace memory backend with markdown in S3; bring your own LLM. Each agent is a workflow for a DAG engine inside a chat completion endpoint. The flexibility-as-feature pattern, packaged.
https://x.com/Everlier/status/2049066541568729453
Jitera launched on Product Hunt after 5 different incarnations. Now: any agent can switch between Jitera's custom agentic loop, Codex, or Claude Code; replace memory backend with markdown in S3; bring your own LLM. Each agent is a workflow for a DAG engine inside a chat completion endpoint. The flexibility-as-feature pattern, packaged.
#27
@grenlouis
https://x.com/grenlouis/status/2049102866967585122
Rebuilt Leon AI core around a deeper agentic loop, better memory, better context injection, more cost-effective execution. Personal AI assistant since 2017, 17.1K stars. Different design center than the new wave: long-term continuity, persistent identity, owner-tied profile data. The "agent that remembers you across years" thesis still has takers.
https://x.com/grenlouis/status/2049102866967585122
Rebuilt Leon AI core around a deeper agentic loop, better memory, better context injection, more cost-effective execution. Personal AI assistant since 2017, 17.1K stars. Different design center than the new wave: long-term continuity, persistent identity, owner-tied profile data. The "agent that remembers you across years" thesis still has takers.
#28
@tandem_engine
https://x.com/tandem_engine/status/2049197216653275640
"The useful agent loop is not prompt β action. It's signal β evidence β proposal β approval β action β memory." Coding agents handle just "action." Everything else is open. Compact framing for what people actually want from agentic systems beyond auto-completing code.
https://x.com/tandem_engine/status/2049197216653275640
"The useful agent loop is not prompt β action. It's signal β evidence β proposal β approval β action β memory." Coding agents handle just "action." Everything else is open. Compact framing for what people actually want from agentic systems beyond auto-completing code.
#29
@youraipulse
https://x.com/youraipulse/status/2049211068447183008
Autoresearch loop pitched as a daily X-content engine: scrape trends, get reports on top performers, reply for engagement, compose content, post on time, grow. Whether or not this specific implementation works, the marketing-side adoption of autoresearch as a primitive is happening fast.
https://x.com/youraipulse/status/2049211068447183008
Autoresearch loop pitched as a daily X-content engine: scrape trends, get reports on top performers, reply for engagement, compose content, post on time, grow. Whether or not this specific implementation works, the marketing-side adoption of autoresearch as a primitive is happening fast.
π‘ Eco Products Radar
Eco Products Radar
Tools, frameworks, and projects mentioned 3+ times in the autoresearch / agentic-loop conversation today:
pi-autoresearch β Pi's autoresearch framework, now on npm with indefinite-runtime support
Karpathy autoresearch β the original primitive, still the reference point in nearly every keeper
Codex / Codex CLI β the most-cited harness people are running autoresearch on
Claude Code β second-most-cited harness, often paired with autoresearch orchestrators
gnhf β open-source tool that wraps autoresearch for any project, 1K+ GitHub stars
OpenClaw β appears in DataClaw and as one of the harnesses agents target
Hermes / Hermes Agent β agent layer in DataClaw, also referenced as the harness running autoresearch loops
GBrain β memory and process structure layer combined with AutoResearch in DataClaw
MCP β connector standard underneath browser-trace, arxiv-radar, and other autoresearch input pipes
Polymarket β the production autoresearch trading target on multiple keepers
Gemini 3.1 Pro β the model running quantitative-finance autoresearch on Google Colab
DeepSeek V4 β both the subject of autoresearch (sglang tuning) and a routing target
GitHub β distribution channel for almost every autoresearch project mentioned
arxiv-radar β MCP server giving agents direct paper access for research loops
browser-trace β derekmeegan's CDP-logging skill explicitly aimed at autoresearch loops
Tools, frameworks, and projects mentioned 3+ times in the autoresearch / agentic-loop conversation today:
pi-autoresearch β Pi's autoresearch framework, now on npm with indefinite-runtime support
Karpathy autoresearch β the original primitive, still the reference point in nearly every keeper
Codex / Codex CLI β the most-cited harness people are running autoresearch on
Claude Code β second-most-cited harness, often paired with autoresearch orchestrators
gnhf β open-source tool that wraps autoresearch for any project, 1K+ GitHub stars
OpenClaw β appears in DataClaw and as one of the harnesses agents target
Hermes / Hermes Agent β agent layer in DataClaw, also referenced as the harness running autoresearch loops
GBrain β memory and process structure layer combined with AutoResearch in DataClaw
MCP β connector standard underneath browser-trace, arxiv-radar, and other autoresearch input pipes
Polymarket β the production autoresearch trading target on multiple keepers
Gemini 3.1 Pro β the model running quantitative-finance autoresearch on Google Colab
DeepSeek V4 β both the subject of autoresearch (sglang tuning) and a routing target
GitHub β distribution channel for almost every autoresearch project mentioned
arxiv-radar β MCP server giving agents direct paper access for research loops
browser-trace β derekmeegan's CDP-logging skill explicitly aimed at autoresearch loops
Comments