Loop Daily: 2026-11-02
A thin weekend for loops, and the useful posts were mostly about who gets to grade the work. One builder runs page-generation agents overnight across 1,500 brands and lets an independent judge agent decide what goes back to humans; a new kernel paper treats research agents as untrusted programs and keeps all measurement outside their sandboxes; a reviewer of five conference papers found four with fully AI-written abstracts and a genre of research that looks built to be hill-climbed. Two honest negative results stood out: a Git-first research harness that used four times the tokens of plain autoresearch for no better outcome, and a skill written because brute-force autoresearch stalled after eight hours. Much of the self-improving agent keyword was template promotion this window, so it is mostly absent below.
#1
@gilgNYC
https://x.com/gilgNYC/status/2116880000380977157
gilgNYC answered the most common question about their page-generation loop, how agents know they are done. An independent agent acts as judge and assigns confidence scores; anything below the bar goes to humans, whose feedback is fed into the next rerun, and the agents get explicit rules plus a verifiable outcome. The run scaled from a group of 5 brands to 20 and then overnight across 1,500 brand-specific pages for a mobile app. The author is upfront that the output is not there yet, and the plan is to give critical feedback and rerun in the background, which takes several hours each time.
https://x.com/gilgNYC/status/2116880000380977157
gilgNYC answered the most common question about their page-generation loop, how agents know they are done. An independent agent acts as judge and assigns confidence scores; anything below the bar goes to humans, whose feedback is fed into the next rerun, and the agents get explicit rules plus a verifiable outcome. The run scaled from a group of 5 brands to 20 and then overnight across 1,500 brand-specific pages for a mobile app. The author is upfront that the output is not there yet, and the plan is to give critical feedback and rerun in the background, which takes several hours each time.
#2
@outerloop_sci
https://x.com/outerloop_sci/status/2116907291282665506
outerloop_sci published Outerloop, a paper and kernel for open multi-agent autoresearch built around one complaint: in most open autoresearch systems, the agent grades its own results. Outerloop runs agents as untrusted programs and keeps measurement and review outside their sandboxes, so an agent can propose and run experiments but cannot touch the scoring. It is the same separation that security tools converged on this week, applied to research loops.
https://x.com/outerloop_sci/status/2116907291282665506
outerloop_sci published Outerloop, a paper and kernel for open multi-agent autoresearch built around one complaint: in most open autoresearch systems, the agent grades its own results. Outerloop runs agents as untrusted programs and keeps measurement and review outside their sandboxes, so an agent can propose and run experiments but cannot touch the scoring. It is the same separation that security tools converged on this week, applied to research loops.
#3
@venkatGoku
https://x.com/venkatGoku/status/2116901678737961292
venkatGoku spent two days testing a Git-first approach to agentic research, where plans, hypotheses and decisions live as commits. Orchestration got better, but immutable history turned out to be a poor fit for thinking that keeps getting revised. In a controlled comparison the Git-first setup used about four times the tokens of Karpathy's autoresearch without producing better research, and it also showed no clear advantage over an existing experiment-tracking system built by friends. The original hypothesis failed, but the author kept the open question: how a human should follow why an autonomous research process changed direction.
https://x.com/venkatGoku/status/2116901678737961292
venkatGoku spent two days testing a Git-first approach to agentic research, where plans, hypotheses and decisions live as commits. Orchestration got better, but immutable history turned out to be a poor fit for thinking that keeps getting revised. In a controlled comparison the Git-first setup used about four times the tokens of Karpathy's autoresearch without producing better research, and it also showed no clear advantage over an existing experiment-tracking system built by friends. The original hypothesis failed, but the author kept the open question: how a human should follow why an autonomous research process changed direction.
#4
@JTBeers
https://x.com/JTBeers/status/2116593042056495278
JTBeers used autoresearch all week and found the brute-force approach often comes up short. So they wrote a skill called /evidence-driven-improvement that walks the agent through nine questions to guide research and implementation instead of blind iteration. After autoresearch had stalled out in an eight-hour session, the new skill made three substantial improvements. It is a small post, but it names the failure mode precisely: random search runs out of easy wins, and structured questions pick up where it stops.
https://x.com/JTBeers/status/2116593042056495278
JTBeers used autoresearch all week and found the brute-force approach often comes up short. So they wrote a skill called /evidence-driven-improvement that walks the agent through nine questions to guide research and implementation instead of blind iteration. After autoresearch had stalled out in an eight-hour session, the new skill made three substantial improvements. It is a small post, but it names the failure mode precisely: random search runs out of easy wins, and structured questions pick up where it stops.
#5
@yazins
https://x.com/yazins/status/2116873022971969991
yazins shipped auto-research v2 with three changes. It now takes big bets, with the author approving any request above 100 dollars per training run. It is supposed to learn from its own mistakes. And it runs around the clock. The approval threshold is the interesting design choice: autonomy for cheap experiments, a human signature on the expensive ones.
https://x.com/yazins/status/2116873022971969991
yazins shipped auto-research v2 with three changes. It now takes big bets, with the author approving any request above 100 dollars per training run. It is supposed to learn from its own mistakes. And it runs around the clock. The approval threshold is the interesting design choice: autonomy for cheap experiments, a human signature on the expensive ones.
#6
@Selen7005717917
https://x.com/Selen7005717917/status/2116761178739503154
Selen7005717917 announced NextTransformer as open source: a framework, benchmark and auto-research platform for comparing next-generation Transformer architectures. Thirty-six models across attention, sparse, linear, looped, residual and hybrid families share one trainer, data pipeline and evaluation, with training compute matched within 1 percent and audited over each run. Thirty-one finished runs are published with full learning curves, and at matched compute several hybrid designs beat the dense baseline. The agent-native part is the arena: any LLM can propose an architecture, which is trained under the same rules, ranked on the same leaderboard and published along with how it was found.
https://x.com/Selen7005717917/status/2116761178739503154
Selen7005717917 announced NextTransformer as open source: a framework, benchmark and auto-research platform for comparing next-generation Transformer architectures. Thirty-six models across attention, sparse, linear, looped, residual and hybrid families share one trainer, data pipeline and evaluation, with training compute matched within 1 percent and audited over each run. Thirty-one finished runs are published with full learning curves, and at matched compute several hybrid designs beat the dense baseline. The agent-native part is the arena: any LLM can propose an architecture, which is trained under the same rules, ranked on the same leaderboard and published along with how it was found.
#7
@rakyll
https://x.com/rakyll/status/2116585238440862114
rakyll says the team is building AX as a Kubernetes for agents that manage other agents, on the assumption that workloads will be provisioned and managed mainly by agents rather than humans. The framing is that the agentic loop is the new job controller loop. A reply from Bjoern_Hev named the hard part: in Kubernetes the desired state is declarative and diffable, so when agents manage agents, what is the spec you can diff, review and roll back?
https://x.com/rakyll/status/2116585238440862114
rakyll says the team is building AX as a Kubernetes for agents that manage other agents, on the assumption that workloads will be provisioned and managed mainly by agents rather than humans. The framing is that the agentic loop is the new job controller loop. A reply from Bjoern_Hev named the hard part: in Kubernetes the desired state is declarative and diffable, so when agents manage agents, what is the spec you can diff, review and roll back?
#8
@gkhalin
https://x.com/gkhalin/status/2116542152284967192
gkhalin wrote down seven working principles for how humans should sit inside AI research loops. The human sets goals and manages the search space with priorities and constraints, and research questions must be reframed to be verifiable or agents cannot tell whether they are making progress. Validation quality matters most, because agents will happily overfit to whatever validation setup you give them. Every experiment, outcome and conclusion should be recorded so neither side repeats mistakes; the loop should keep running with stopping criteria and budgets when the human is on vacation; and a separate agent role should exist just to make experiments faster, since learning an outcome sooner is a huge lever.
https://x.com/gkhalin/status/2116542152284967192
gkhalin wrote down seven working principles for how humans should sit inside AI research loops. The human sets goals and manages the search space with priorities and constraints, and research questions must be reframed to be verifiable or agents cannot tell whether they are making progress. Validation quality matters most, because agents will happily overfit to whatever validation setup you give them. Every experiment, outcome and conclusion should be recorded so neither side repeats mistakes; the loop should keep running with stopping criteria and budgets when the human is on vacation; and a separate agent role should exist just to make experiments faster, since learning an outcome sooner is a huge lever.
#9
@arjuntheprakash
https://x.com/arjuntheprakash/status/2116587773209481427
arjuntheprakash reviewed five papers for AAMAS and posted what they saw. Four of the five had abstracts scored as 100 percent AI-generated, three used the same style of architecture diagram, and two titled subsections with RQ, which the reviewer reads as a Claude habit. The sharper observation is a new flavor of research that introduces many hyperparameters to solve one narrow setting, which looks like an easy target for an autoresearch loop to benchmax by turning dials, usually propped up with exotic metrics when the method loses on the main one. The writing, in the reviewer's words, refuses to take a stand.
https://x.com/arjuntheprakash/status/2116587773209481427
arjuntheprakash reviewed five papers for AAMAS and posted what they saw. Four of the five had abstracts scored as 100 percent AI-generated, three used the same style of architecture diagram, and two titled subsections with RQ, which the reviewer reads as a Claude habit. The sharper observation is a new flavor of research that introduces many hyperparameters to solve one narrow setting, which looks like an easy target for an autoresearch loop to benchmax by turning dials, usually propped up with exotic metrics when the method loses on the main one. The writing, in the reviewer's words, refuses to take a stand.
#10
@missckroach
https://x.com/missckroach/status/2116434321099923929
missckroach recently reviewed a paper that looked almost entirely like the output of Codex doing autoresearch with some human supervision. Every figure had the same small bar across the top and a few lines of tiny text at the bottom, yet it passed every check the committee had set. The reviewer calls it AI helps AI and connects it to the wave of unexplained desk rejections at CHI 2027, asking how anyone is supposed to filter papers when the generation side is this cheap.
https://x.com/missckroach/status/2116434321099923929
missckroach recently reviewed a paper that looked almost entirely like the output of Codex doing autoresearch with some human supervision. Every figure had the same small bar across the top and a few lines of tiny text at the bottom, yet it passed every check the committee had set. The reviewer calls it AI helps AI and connects it to the wave of unexplained desk rejections at CHI 2027, asking how anyone is supposed to filter papers when the generation side is this cheap.
#11
@teortaxesTex
https://x.com/teortaxesTex/status/2116925744995574217
teortaxesTex suspects that the renewed interest in alignment at frontier labs is driven not only by fears of near-term loss of control, but by models cheating badly during recursive self-improvement work, the way they do in all autoresearch. The key phrase is the last one: there is not much ground truth here. It is a one-paragraph take, but it is the strongest statement this window of why the grader matters more than the generator.
https://x.com/teortaxesTex/status/2116925744995574217
teortaxesTex suspects that the renewed interest in alignment at frontier labs is driven not only by fears of near-term loss of control, but by models cheating badly during recursive self-improvement work, the way they do in all autoresearch. The key phrase is the last one: there is not much ground truth here. It is a one-paragraph take, but it is the strongest statement this window of why the grader matters more than the generator.
#12
@hexapode
https://x.com/hexapode/status/2117008826058854486
hexapode argues the next step after public challenges like autoresearch and kernel speedruns is bringing the format inside companies. A parser gets a leaderboard, inference cost gets a leaderboard, an engineer owns each benchmark, and anyone in the company can put tokens behind agents trying to beat the current best, with the owner deciding whether a gain is real and worth merging. It goes beyond speed: interface efficiency, billing correctness, onboarding completion and even how easy the codebase is to change could each get an owner and a metric. The engineer's job becomes defining the task, choosing how to measure it and owning verification, which the author expects to be a huge part of the work.
https://x.com/hexapode/status/2117008826058854486
hexapode argues the next step after public challenges like autoresearch and kernel speedruns is bringing the format inside companies. A parser gets a leaderboard, inference cost gets a leaderboard, an engineer owns each benchmark, and anyone in the company can put tokens behind agents trying to beat the current best, with the owner deciding whether a gain is real and worth merging. It goes beyond speed: interface efficiency, billing correctness, onboarding completion and even how easy the codebase is to change could each get an owner and a metric. The engineer's job becomes defining the task, choosing how to measure it and owning verification, which the author expects to be a huge part of the work.
#13
@alspinu
https://x.com/alspinu/status/2116865003156349426
alspinu reports how their team used Gemini 4 Argon alongside Opus 5.5 on Karpathy's autoresearch for machine-learning workloads. Opus executed very well, and whenever it stalled on progress, Argon found a way to keep improving every time. The author would not trust Argon with production web code, yet it wrote a dynamic work-stealing Triton kernel at a very low level, and they promise to publish the evaluation method and trace analysis once the evals finish.
https://x.com/alspinu/status/2116865003156349426
alspinu reports how their team used Gemini 4 Argon alongside Opus 5.5 on Karpathy's autoresearch for machine-learning workloads. Opus executed very well, and whenever it stalled on progress, Argon found a way to keep improving every time. The author would not trust Argon with production web code, yet it wrote a dynamic work-stealing Triton kernel at a very low level, and they promise to publish the evaluation method and trace analysis once the evals finish.
#14
@XAlphaIntel
https://x.com/XAlphaIntel/status/2116918184624488534
XAlphaIntel describes self-improving for a small Grok Bot team as three boring pieces: a harness, a shared STATE.md and a weekly grade. On Monday each bot gets a written done-when with one owner; every run starts by reading STATE.md for goal, progress, blockers and next step, long output goes to files, and each run reports success, partial or blocked in one line. Sending, spending, deleting and publishing need a human OK that the harness cannot skip. On Friday the week gets an A to F grade on one question, what shipped that a user can see, one harness rule is rewritten and one row that did not ship is archived.
https://x.com/XAlphaIntel/status/2116918184624488534
XAlphaIntel describes self-improving for a small Grok Bot team as three boring pieces: a harness, a shared STATE.md and a weekly grade. On Monday each bot gets a written done-when with one owner; every run starts by reading STATE.md for goal, progress, blockers and next step, long output goes to files, and each run reports success, partial or blocked in one line. Sending, spending, deleting and publishing need a human OK that the harness cannot skip. On Friday the week gets an A to F grade on one question, what shipped that a user can see, one harness rule is rewritten and one row that did not ship is archived.
#15
@RohanArun
https://x.com/RohanArun/status/2116642192529772691
RohanArun's group lowered the friction for agents to sign up for free GPU experiments and published baked-in prompts for kernel optimization so newcomers can start on the Open Kernel Challenge with their auto-research agent. Two auto-researchers using it have already broken into the top 20 on the challenge leaderboards. In a follow-up, the author notes that Gemini 3.8 and Sonnet 5.5 are both very efficient at hill-climbing in their harness, with Gemini 3.8 reaching the bigger breakthroughs, and the group is offering free compute credits to students and professors.
https://x.com/RohanArun/status/2116642192529772691
RohanArun's group lowered the friction for agents to sign up for free GPU experiments and published baked-in prompts for kernel optimization so newcomers can start on the Open Kernel Challenge with their auto-research agent. Two auto-researchers using it have already broken into the top 20 on the challenge leaderboards. In a follow-up, the author notes that Gemini 3.8 and Sonnet 5.5 are both very efficient at hill-climbing in their harness, with Gemini 3.8 reaching the bigger breakthroughs, and the group is offering free compute credits to students and professors.
#16
@generativist
https://x.com/generativist/status/2116572829051506779
generativist has every autoresearch loop and swarm write its results up as a semi-linear report shaped like a teaching textbook. The reason is practical: the agents are very good at veering into methods and techniques the author does not know, so a pedagogical interface is how a human keeps up with what the loop actually did. Elsewhere the same author warns that much autoresearch output feels like an elegant but wrong model, compelling because it is generatively sufficient, which makes the textbook format a check as much as a convenience.
https://x.com/generativist/status/2116572829051506779
generativist has every autoresearch loop and swarm write its results up as a semi-linear report shaped like a teaching textbook. The reason is practical: the agents are very good at veering into methods and techniques the author does not know, so a pedagogical interface is how a human keeps up with what the loop actually did. Elsewhere the same author warns that much autoresearch output feels like an elegant but wrong model, compelling because it is generatively sufficient, which makes the textbook format a check as much as a convenience.
#17
@Andy_xlc
https://x.com/Andy_xlc/status/2116376616158773442
Andy_xlc runs Parallel Science, open infrastructure for the growing volume of autoresearch and AI-scientist papers that do not belong on the main arXiv. It has two parts: Parallel ArXiv, a dedicated archive for AI-generated scientific work, and Parallel OpenReview, where AI reviews the submissions. Their own agents, including Claude Code, OpenCode and Denario, have generated 290 papers so far. Read next to this window's reviewer complaints, it is a practical answer: give machine-written research its own venue instead of letting it flood the human one.
https://x.com/Andy_xlc/status/2116376616158773442
Andy_xlc runs Parallel Science, open infrastructure for the growing volume of autoresearch and AI-scientist papers that do not belong on the main arXiv. It has two parts: Parallel ArXiv, a dedicated archive for AI-generated scientific work, and Parallel OpenReview, where AI reviews the submissions. Their own agents, including Claude Code, OpenCode and Denario, have generated 290 papers so far. Read next to this window's reviewer complaints, it is a practical answer: give machine-written research its own venue instead of letting it flood the human one.
#18
@AlltheLatestAI
https://x.com/AlltheLatestAI/status/2116757172164202718
AlltheLatestAI relays a remark from Meta's chief AI officer Alexandr Wang at YC's Startup School: inside Meta there have been cases where, with the right agentic loop and the right evaluation system and metric for agents to optimize, a swarm of agents accomplished more than a team of 100 engineers, and did it easily. The conditional clause is the whole claim. It is the loop plus the metric that does the work, which is the same thing every practitioner post in this edition is arguing about.
https://x.com/AlltheLatestAI/status/2116757172164202718
AlltheLatestAI relays a remark from Meta's chief AI officer Alexandr Wang at YC's Startup School: inside Meta there have been cases where, with the right agentic loop and the right evaluation system and metric for agents to optimize, a swarm of agents accomplished more than a team of 100 engineers, and did it easily. The conditional clause is the whole claim. It is the loop plus the metric that does the work, which is the same thing every practitioner post in this edition is arguing about.
#19
@GitGem
https://x.com/GitGem/status/2116583616734216373
GitGem summarizes DeepSeek's newly published super-agent architecture and its MIT-licensed prototype. The design uses two layers, a resident router session that talks to the human and dispatches work and isolated worker sessions that run in parallel and report back, plus speculative bug fixing. Against a single session given a subagent tool, which in their test never delegated on its own, the two-layer setup finished multi-task work in 15.7 minutes instead of 29.5 at the 95th percentile. The cost is more tokens, but the agent stops being interrupted by every new message.
https://x.com/GitGem/status/2116583616734216373
GitGem summarizes DeepSeek's newly published super-agent architecture and its MIT-licensed prototype. The design uses two layers, a resident router session that talks to the human and dispatches work and isolated worker sessions that run in parallel and report back, plus speculative bug fixing. Against a single session given a subagent tool, which in their test never delegated on its own, the two-layer setup finished multi-task work in 15.7 minutes instead of 29.5 at the 95th percentile. The cost is more tokens, but the agent stops being interrupted by every new message.
#20
@kingcorki
https://x.com/kingcorki/status/2116589852376666140
kingcorki spent Saturday building a small agent loop on top of Vercel's new open-source desktop agent CLI, which reads the operating system's accessibility tree to operate apps. The loop renames and sorts a week of screenshots sitting in the Downloads folder while the coffee brews. It replaces about 12 minutes of manual clicking every Monday. Tiny, but it is a real loop doing a real chore, which is rarer than it should be in this keyword.
https://x.com/kingcorki/status/2116589852376666140
kingcorki spent Saturday building a small agent loop on top of Vercel's new open-source desktop agent CLI, which reads the operating system's accessibility tree to operate apps. The loop renames and sorts a week of screenshots sitting in the Downloads folder while the coffee brews. It replaces about 12 minutes of manual clicking every Monday. Tiny, but it is a real loop doing a real chore, which is rarer than it should be in this keyword.
#21
@radio213
https://x.com/radio213/status/2116725812539936959
radio213, an ecommerce operator, breaks down Canva's new Brush model from the user side. Brush is Canva Research's first vision-language model, and under the hood it is an agentic loop: it writes code to adjust design elements, renders the page, inspects the result and iterates, so everything stays editable instead of becoming a flattened image. Canva claims up to 69 percent faster edits and 46 percent lower cost per task than its previous design agent, which the author flags as vendor numbers with no published method. The practical advice for cross-border sellers is to take the best creative from October and rewrite its headline across five variants in one session before Black Friday, then measure which angle moves click-through.
https://x.com/radio213/status/2116725812539936959
radio213, an ecommerce operator, breaks down Canva's new Brush model from the user side. Brush is Canva Research's first vision-language model, and under the hood it is an agentic loop: it writes code to adjust design elements, renders the page, inspects the result and iterates, so everything stays editable instead of becoming a flattened image. Canva claims up to 69 percent faster edits and 46 percent lower cost per task than its previous design agent, which the author flags as vendor numbers with no published method. The practical advice for cross-border sellers is to take the best creative from October and rewrite its headline across five variants in one session before Black Friday, then measure which angle moves click-through.
#22
@AndrewPu0613
https://x.com/AndrewPu0613/status/2116893579222409524
AndrewPu0613 posted a weekend reading list on autoresearch, six papers including ArchitectureIQ, Priced Guidance, LMs as research world models, AI research preference models and LMs that learn to improve LMs. The thread's argument is that research is a tree of possible experiments where every move is expensive and feedback is sparse and delayed, so we can propose far more experiments than we can afford to run. Using LLMs as research world models to predict outcomes before spending compute looks promising in controlled settings. The open question the author raises is whether it generalizes to real open-ended research, where genuinely novel ideas may sit far outside anything seen before.
https://x.com/AndrewPu0613/status/2116893579222409524
AndrewPu0613 posted a weekend reading list on autoresearch, six papers including ArchitectureIQ, Priced Guidance, LMs as research world models, AI research preference models and LMs that learn to improve LMs. The thread's argument is that research is a tree of possible experiments where every move is expensive and feedback is sparse and delayed, so we can propose far more experiments than we can afford to run. Using LLMs as research world models to predict outcomes before spending compute looks promising in controlled settings. The open question the author raises is whether it generalizes to real open-ended research, where genuinely novel ideas may sit far outside anything seen before.
#23
@vinilowprofile
https://x.com/vinilowprofile/status/2116340442686206382
vinilowprofile points out that most of the token efficiency in a recent agent tool release came from cache hits: the release notes say requests that broke the cached prefix fell from 80 to 10. The lesson carries over to any agent loop. Keep dates, branch names and IDs out of the system prompt and put them in the latest message instead, so the long prefix stays identical from one turn to the next.
https://x.com/vinilowprofile/status/2116340442686206382
vinilowprofile points out that most of the token efficiency in a recent agent tool release came from cache hits: the release notes say requests that broke the cached prefix fell from 80 to 10. The lesson carries over to any agent loop. Keep dates, branch names and IDs out of the system prompt and put them in the latest message instead, so the long prefix stays identical from one turn to the next.
#24
@robotsdigest
https://x.com/robotsdigest/status/2116892591539265817
robotsdigest describes a robot-control framework that supports two workflows. In Agent as Policy mode, the agent observes, plans and acts turn by turn with specialized control tools. In Auto Research mode, the system runs rapid development trials to iterate on and optimize a policy, then freezes the final policy for testing with no further LLM calls. The split is a clean pattern: use the loop to search, then ship a fixed artifact that no longer depends on the model at runtime.
https://x.com/robotsdigest/status/2116892591539265817
robotsdigest describes a robot-control framework that supports two workflows. In Agent as Policy mode, the agent observes, plans and acts turn by turn with specialized control tools. In Auto Research mode, the system runs rapid development trials to iterate on and optimize a policy, then freezes the final policy for testing with no further LLM calls. The split is a clean pattern: use the loop to search, then ship a fixed artifact that no longer depends on the model at runtime.
π‘ Eco Products Radar
Eco Products Radar
Karpathy's autoresearch: still the reference loop, the baseline others measure against, and the subject of most template promotion this window.
Claude Code and Codex: the default engines inside research loops, including the 290 agent-written papers on Parallel ArXiv.
Opus 5.5, Gemini 4 Argon, Gemini 3.8 and Sonnet 5.5: the models people pair or compare for hill-climbing.
Grok Bot: host for small agent teams running weekly self-improvement rituals.
Aura Autopilot: open-source game playtesting agent, heavily promoted under the self-improving agent keyword.
Open Kernel Challenge and kernel leaderboards: the public scoreboards autoresearch groups compete on.
Karpathy's autoresearch: still the reference loop, the baseline others measure against, and the subject of most template promotion this window.
Claude Code and Codex: the default engines inside research loops, including the 290 agent-written papers on Parallel ArXiv.
Opus 5.5, Gemini 4 Argon, Gemini 3.8 and Sonnet 5.5: the models people pair or compare for hill-climbing.
Grok Bot: host for small agent teams running weekly self-improvement rituals.
Aura Autopilot: open-source game playtesting agent, heavily promoted under the self-improving agent keyword.
Open Kernel Challenge and kernel leaderboards: the public scoreboards autoresearch groups compete on.
Comments