Loop Daily: 2026-10-28
This window autoresearch went to the leaderboard. On Meta's Open Kernel Challenge one user reached ninth place with a single prompt left running overnight, 16 hours, about 150 H200 experiments and 250 dollars, while another team climbed 110 ranks in a day and reported that human coordinators steering a fleet of cheap agents move more than twice as fast as an AI coordinator alone. The loop is also leaving the lab: an app team put the whole install-to-paywall funnel into an agentic loop and lifted ARPU from 0.45 to 0.74 dollars, a freelancer rebuilt a team of Claude Code agents around one coordinator and tap-to-answer decisions, and Recursive shipped a model in four months with auto-research at almost every stage. The research side kept getting more sober. Auto-Research Arena found plain sequential search ranks second while a tuned TPE optimizer ranks last, and Meta's own report showed a poisoned skill succeeding six times more often inside its agent harness than through the API, because the harness trusts its own skill folder. The common thread: the loop is cheap, and what you let it trust is the expensive part.
#1
@naklecha
https://x.com/naklecha/status/2115172803997483508
naklecha says an open-source autoresearch harness turned out to be great at writing fast kernels: one prompt typed the night before put them at number 9 on Meta's kernel leaderboard. The run was 16 hours fully autonomous, about 150 H200 experiments, ending 1.57 times faster than SGLang's kernels, for roughly 250 dollars of tokens plus compute. The cost line is the story: a leaderboard position that used to require a GPU kernel specialist was bought overnight at the price of a nice dinner. It is also the clearest single data point that the bottleneck in kernel work has moved from writing code to paying for experiments.
https://x.com/naklecha/status/2115172803997483508
naklecha says an open-source autoresearch harness turned out to be great at writing fast kernels: one prompt typed the night before put them at number 9 on Meta's kernel leaderboard. The run was 16 hours fully autonomous, about 150 H200 experiments, ending 1.57 times faster than SGLang's kernels, for roughly 250 dollars of tokens plus compute. The cost line is the story: a leaderboard position that used to require a GPU kernel specialist was bought overnight at the price of a nice dinner. It is also the clearest single data point that the bottleneck in kernel work has moved from writing code to paying for experiments.
#2
@RohanArun
https://x.com/RohanArun/status/2115093010832630019
RohanArun's team climbed 110 ranks across the Open Kernel Challenge leaderboards in a day, posting each before-and-after: prefill from 297,235 to 266,196 microseconds and up 74 ranks, attention decode up 36, while three other boards slipped as more teams joined. The most useful observation is about coordination: human experts steering agents progressed more than twice as fast as an AI coordinator alone, because even frontier models make mistakes experts can fix, the most common being forgetting that kernel optimizations can be combined into the full decode layer. They hill-climbed with a fleet of cheap flash-model agents plus a human coordinator, and opened their GPU queue, pricing experiments in H200 seconds and claiming about 90 percent savings over renting an on-demand instance.
https://x.com/RohanArun/status/2115093010832630019
RohanArun's team climbed 110 ranks across the Open Kernel Challenge leaderboards in a day, posting each before-and-after: prefill from 297,235 to 266,196 microseconds and up 74 ranks, attention decode up 36, while three other boards slipped as more teams joined. The most useful observation is about coordination: human experts steering agents progressed more than twice as fast as an AI coordinator alone, because even frontier models make mistakes experts can fix, the most common being forgetting that kernel optimizations can be combined into the full decode layer. They hill-climbed with a fleet of cheap flash-model agents plus a human coordinator, and opened their GPU queue, pricing experiments in H200 seconds and claiming about 90 percent savings over renting an on-demand instance.
#3
@smgdkngt
https://x.com/smgdkngt/status/2115021751616709054
smgdkngt runs a freelance business with a team of Claude Code agents and rebuilt the setup after reading Anthropic's coordination and subagent-folding pieces. One coordinator session reads every email, routine report and project event, and each client has a standing domain session that keeps its own people, systems, open work and skill file. Every decision reaches the human as one tappable question with a recommended answer first, progress lives on cards updated through an MCP so a fresh session resumes exactly where the last stopped, and deploys, merges and production writes become scripts in an approval queue. Every miss becomes an eval case that a fresh agent replays without the answers, which is the loop doing the improving rather than the human.
https://x.com/smgdkngt/status/2115021751616709054
smgdkngt runs a freelance business with a team of Claude Code agents and rebuilt the setup after reading Anthropic's coordination and subagent-folding pieces. One coordinator session reads every email, routine report and project event, and each client has a standing domain session that keeps its own people, systems, open work and skill file. Every decision reaches the human as one tappable question with a recommended answer first, progress lives on cards updated through an MCP so a fresh session resumes exactly where the last stopped, and deploys, merges and production writes become scripts in an approval queue. Every miss becomes an eval case that a fresh agent replays without the answers, which is the loop doing the improving rather than the human.
#4
@jakemor
https://x.com/jakemor/status/2115179208674775504
jakemor's team is testing an agent that puts the whole mobile funnel, from install to paywall, into an agentic loop driven by the Superwall CLI. Early results are specific: it added a paywall at install because there was none, which first pushed ARPU down from 0.58 to 0.45 dollars, then it hill-climbed a series of onboarding questions to prime that paywall and took ARPU to 0.74. That is a 27 percent overall increase with more revenue up front. It is one of the clearest non-research autoresearch cases yet, because the metric is money and the action space is a real product.
https://x.com/jakemor/status/2115179208674775504
jakemor's team is testing an agent that puts the whole mobile funnel, from install to paywall, into an agentic loop driven by the Superwall CLI. Early results are specific: it added a paywall at install because there was none, which first pushed ARPU down from 0.58 to 0.45 dollars, then it hill-climbed a series of onboarding questions to prime that paywall and took ARPU to 0.74. That is a 27 percent overall increase with more revenue up front. It is one of the clearest non-research autoresearch cases yet, because the metric is money and the action space is a real product.
#5
@altryne
https://x.com/altryne/status/2115069871872729236
altryne shared a clip from Tim Shi of Recursive, who said the company had no pre-training or post-training effort until about a month before shipping its model, four months from scratch overall. Their answer was to run auto-research on almost every stage of the pipeline, from pre-training to RL to inference. It is a lab-scale version of the same claim the kernel leaderboard is making at hobby scale: the loop compresses calendar time when the evaluation is good. The obvious caveat, raised in the replies, is that auto-research collapses when the eval it optimizes drifts from what users hit in production.
https://x.com/altryne/status/2115069871872729236
altryne shared a clip from Tim Shi of Recursive, who said the company had no pre-training or post-training effort until about a month before shipping its model, four months from scratch overall. Their answer was to run auto-research on almost every stage of the pipeline, from pre-training to RL to inference. It is a lab-scale version of the same claim the kernel leaderboard is making at hobby scale: the loop compresses calendar time when the evaluation is good. The obvious caveat, raised in the replies, is that auto-research collapses when the eval it optimizes drifts from what users hit in production.
#6
@AmericaGetBig
https://x.com/AmericaGetBig/status/2115195501775913067
AmericaGetBig summarized the Auto-Research Arena paper from Xi Chen, Henrik Christensen and Adobe Research, which pits 7 research harnesses and 2 LLMs against 9 targets, from model size to inference speed, under a quality floor. The headline result is humbling for clever search: plain sequential search ranks second with both LLMs, while TPE, a method tuned for validation loss, ranks last. Qualifying for a target is scored separately from improving on it, which keeps a harness from looking good just by clearing the floor. It argues that most of the value is in the loop and the evaluation, not in a smarter search algorithm on top.
https://x.com/AmericaGetBig/status/2115195501775913067
AmericaGetBig summarized the Auto-Research Arena paper from Xi Chen, Henrik Christensen and Adobe Research, which pits 7 research harnesses and 2 LLMs against 9 targets, from model size to inference speed, under a quality floor. The headline result is humbling for clever search: plain sequential search ranks second with both LLMs, while TPE, a method tuned for validation loss, ranks last. Qualifying for a target is scored separately from improving on it, which keeps a harness from looking good just by clearing the floor. It argues that most of the value is in the loop and the evaluation, not in a smarter search algorithm on top.
#7
@hanwenjiang1
https://x.com/hanwenjiang1/status/2115137967660597332
hanwenjiang1, one of the Auto-Research Arena authors, highlighted a second result: the multi-target search builds a large idea bank as a side effect, and ideas found in cheap 10-minute runs improve performance when applied to longer-term runs. That matters for anyone paying for compute, because it suggests a two-stage loop, many short exploratory runs to harvest ideas and a few long runs to confirm them. It also pairs well with the kernel crowd's finding that the expensive part is experiments, not code.
https://x.com/hanwenjiang1/status/2115137967660597332
hanwenjiang1, one of the Auto-Research Arena authors, highlighted a second result: the multi-target search builds a large idea bank as a side effect, and ideas found in cheap 10-minute runs improve performance when applied to longer-term runs. That matters for anyone paying for compute, because it suggests a two-stage loop, many short exploratory runs to harvest ideas and a few long runs to confirm them. It also pairs well with the kernel crowd's finding that the expensive part is experiments, not code.
#8
@hammer_mt
https://x.com/hammer_mt/status/2115122380297089078
hammer_mt is running multiplayer autoresearch, where several models work the same task and can see each other's log of attempts, and reports Gemini 4 Argon set every frontier. Its round-3 breakthrough from 0.625 to 0.938 opened by citing round 2's failures: Fable and Sol had both loosened a long-sentence rule and dropped the score, so Argon kept that rule and changed something else. In round 4 Fable and Sol scored 0.938 and 0.875 by building on Argon's skill. It is a small but concrete demonstration that a shared failure log is worth as much as a shared success, the same intuition behind fork-and-flush.
https://x.com/hammer_mt/status/2115122380297089078
hammer_mt is running multiplayer autoresearch, where several models work the same task and can see each other's log of attempts, and reports Gemini 4 Argon set every frontier. Its round-3 breakthrough from 0.625 to 0.938 opened by citing round 2's failures: Fable and Sol had both loosened a long-sentence rule and dropped the score, so Argon kept that rule and changed something else. In round 4 Fable and Sol scored 0.938 and 0.875 by building on Argon's skill. It is a small but concrete demonstration that a shared failure log is worth as much as a shared success, the same intuition behind fork-and-flush.
#9
@miu21590
https://x.com/miu21590/status/2114924456854352098
miu21590 has been testing Octopus OS, an open-source agent system designed so agents can change their own harness as part of ongoing work and restart the system when needed. A trigger that fires after an hour of inactivity looked into what went wrong during use, patched the source code, tested the fix and applied it, here a bug in voice mode, without stopping the agent. Tools, context, external event sources and the conditions for starting or continuing work are all editable, with no hardcoded agent loop in charge. The desktop app is in research preview with the trigger on by default in custom setup, which makes it one of the more literal self-improving-harness experiments in the wild.
https://x.com/miu21590/status/2114924456854352098
miu21590 has been testing Octopus OS, an open-source agent system designed so agents can change their own harness as part of ongoing work and restart the system when needed. A trigger that fires after an hour of inactivity looked into what went wrong during use, patched the source code, tested the fix and applied it, here a bug in voice mode, without stopping the agent. Tools, context, external event sources and the conditions for starting or continuing work are all editable, with no hardcoded agent loop in charge. The desktop app is in research preview with the trigger on by default in custom setup, which makes it one of the more literal self-improving-harness experiments in the wild.
#10
@kbhaskar_95
https://x.com/kbhaskar_95/status/2115171928520397136
kbhaskar_95 dug into a detail in Meta's own model report: on the SkillInject benchmark, where a normal task comes with a skill file carrying an injected instruction, the same model attempted the planted instruction in 0.6 percent of runs through the API and 3.7 percent inside Muse, Meta's agent harness. Meta's explanation is that Muse is built to load skills and follow them, so a malicious skill in its skill directory arrives through a channel the harness treats as trusted. The harness is not always worse, scoring 10 percent against the API's 17.5 on browser prompt injection thanks to system-level mitigations. The practical advice is the useful part: if your agent loads skills from a folder, put the test skill in that folder and run your real loader.
https://x.com/kbhaskar_95/status/2115171928520397136
kbhaskar_95 dug into a detail in Meta's own model report: on the SkillInject benchmark, where a normal task comes with a skill file carrying an injected instruction, the same model attempted the planted instruction in 0.6 percent of runs through the API and 3.7 percent inside Muse, Meta's agent harness. Meta's explanation is that Muse is built to load skills and follow them, so a malicious skill in its skill directory arrives through a channel the harness treats as trusted. The harness is not always worse, scoring 10 percent against the API's 17.5 on browser prompt injection thanks to system-level mitigations. The practical advice is the useful part: if your agent loads skills from a folder, put the test skill in that folder and run your real loader.
#11
@RobertTLange
https://x.com/RobertTLange/status/2115110100616855900
RobertTLange highlighted Fork-and-Flush, a paper that frames autoresearch as idea descent that can get stuck in idea basins the way gradient descent gets stuck in local optima. The method forks the agent into K branches that keep the workspace but start with an empty chat history, runs each for part of the budget, keeps the best and continues; a single long run and best-of-N are the two extremes. At equal compute it beats both across research and engineering tasks with less run-to-run variance, and forking and flushing each help but work best together. The insight is that dropping the chat forces a branch to rediscover intent from the files alone, which is what lets it land in a new basin.
https://x.com/RobertTLange/status/2115110100616855900
RobertTLange highlighted Fork-and-Flush, a paper that frames autoresearch as idea descent that can get stuck in idea basins the way gradient descent gets stuck in local optima. The method forks the agent into K branches that keep the workspace but start with an empty chat history, runs each for part of the budget, keeps the best and continues; a single long run and best-of-N are the two extremes. At equal compute it beats both across research and engineering tasks with less run-to-run variance, and forking and flushing each help but work best together. The insight is that dropping the chat forces a branch to rediscover intent from the files alone, which is what lets it land in a new basin.
#12
@daibond_alpha
https://x.com/daibond_alpha/status/2115170739938795779
daibond_alpha's group gave their Kinex physical agent a self-improving architecture with recursive self-improvement ability and report it solved all tasks in the RoboCasa household manipulation benchmark with only a handful of experiences. The post links the robot, a physical RSI write-up and the author's thoughts on self-improving agents, credited to a team of students. Details are thin in the tweet itself, but it is one of the few claims this window that moves the self-improving loop off the screen and onto a robot, where every failed experience is expensive.
https://x.com/daibond_alpha/status/2115170739938795779
daibond_alpha's group gave their Kinex physical agent a self-improving architecture with recursive self-improvement ability and report it solved all tasks in the RoboCasa household manipulation benchmark with only a handful of experiences. The post links the robot, a physical RSI write-up and the author's thoughts on self-improving agents, credited to a team of students. Details are thin in the tweet itself, but it is one of the few claims this window that moves the self-improving loop off the screen and onto a robot, where every failed experience is expensive.
#13
@joelhooks
https://x.com/joelhooks/status/2114818780111286403
joelhooks says an autonomous agent loop now negotiates enterprise sales daily in the 2,000 to 75,000 dollar range. It answers initial questions, handles zip and procurement onboarding, negotiates contracts and discounts, and covers fulfillment and post-sale support. It is a short post with no numbers on win rate, but it is a striking example of how far the loop has moved from overnight experiments into a revenue-bearing workflow with a human counterparty on the other side.
https://x.com/joelhooks/status/2114818780111286403
joelhooks says an autonomous agent loop now negotiates enterprise sales daily in the 2,000 to 75,000 dollar range. It answers initial questions, handles zip and procurement onboarding, negotiates contracts and discounts, and covers fulfillment and post-sale support. It is a short post with no numbers on win rate, but it is a striking example of how far the loop has moved from overnight experiments into a revenue-bearing workflow with a human counterparty on the other side.
#14
@kumo0x
https://x.com/kumo0x/status/2114862902368768112
kumo0x summarized a second-brain guide built on three layers, Obsidian as plain markdown source of truth, Hermes Agent reading folders and running scheduled jobs, and MiniMax M3 as the model that can read a whole vault at once, run as a loop rather than one-off chat. The author reports about 41 tags in a fixed schema with the right primary tag on the first try about 90 percent of the time versus about 60 on a 200K-context model, and full vault lints of 30-plus tool calls staying coherent where most models drift around call 8 or 9. The permission rule is the transferable part: humans write freely in Inbox, Daily and Reading, and the agent only writes into Projects and Reviews. The added advice is to put the vault under git first so every agent edit can be diffed and undone.
https://x.com/kumo0x/status/2114862902368768112
kumo0x summarized a second-brain guide built on three layers, Obsidian as plain markdown source of truth, Hermes Agent reading folders and running scheduled jobs, and MiniMax M3 as the model that can read a whole vault at once, run as a loop rather than one-off chat. The author reports about 41 tags in a fixed schema with the right primary tag on the first try about 90 percent of the time versus about 60 on a 200K-context model, and full vault lints of 30-plus tool calls staying coherent where most models drift around call 8 or 9. The permission rule is the transferable part: humans write freely in Inbox, Daily and Reading, and the agent only writes into Projects and Reviews. The added advice is to put the vault under git first so every agent edit can be diffed and undone.
#15
@muratnakmuay
https://x.com/muratnakmuay/status/2114613206169653541
muratnakmuay summarized a blog timeline of frontier API price cuts of up to 80 percent in two months, from GPT 5.6 Luna dropping to 0.20 and 1.20 dollars per million tokens to Opus 5.5 launching at 4 and 20, and argued anyone who priced an agent loop in summer should redo the math at September prices. The detail aimed at loop builders is that cache reads are where long sessions spend most tokens, so cache price cuts matter more than list prices. The sharper point is about plans rather than APIs: the blog author's Codex quota fell to a fifth after the model got about four times more efficient, so the effective cost stayed flat. The prediction is that subscription allowances will shrink before API prices rise.
https://x.com/muratnakmuay/status/2114613206169653541
muratnakmuay summarized a blog timeline of frontier API price cuts of up to 80 percent in two months, from GPT 5.6 Luna dropping to 0.20 and 1.20 dollars per million tokens to Opus 5.5 launching at 4 and 20, and argued anyone who priced an agent loop in summer should redo the math at September prices. The detail aimed at loop builders is that cache reads are where long sessions spend most tokens, so cache price cuts matter more than list prices. The sharper point is about plans rather than APIs: the blog author's Codex quota fell to a fifth after the model got about four times more efficient, so the effective cost stayed flat. The prediction is that subscription allowances will shrink before API prices rise.
#16
@my_cat_can_code
https://x.com/my_cat_can_code/status/2114859255807578509
my_cat_can_code is running Investment Arena: 72 AI agents, 18 models with 4 trading strategies each, every one handed 120,000 dollars of paper money and a week to trade stocks, crypto and Polymarket, live and in public. Beyond a chat window they get dozens of information sources, historical market data, live news and CPU and GPU sandboxes to build and train their own market models, so a thesis can be tested, traded, and revised after it fails. It follows the same team's RSI Arena, and the framing is a full week of nonstop auto research meeting a market that keeps moving. It is paper money, but it is one of the few loops this week with a live, non-gameable scoreboard.
https://x.com/my_cat_can_code/status/2114859255807578509
my_cat_can_code is running Investment Arena: 72 AI agents, 18 models with 4 trading strategies each, every one handed 120,000 dollars of paper money and a week to trade stocks, crypto and Polymarket, live and in public. Beyond a chat window they get dozens of information sources, historical market data, live news and CPU and GPU sandboxes to build and train their own market models, so a thesis can be tested, traded, and revised after it fails. It follows the same team's RSI Arena, and the framing is a full week of nonstop auto research meeting a market that keeps moving. It is paper money, but it is one of the few loops this week with a live, non-gameable scoreboard.
#17
@aaronspradlin
https://x.com/aaronspradlin/status/2114909040283472383
aaronspradlin described wiring a multi-agent Grok Bot loop to do 24/7 math and physics research with one hard rule, the bots were banned from bluffing. Grok Bot proposes theorems and then verifies them exhaustively in Python, all 59,829 states with zero errors reported. The design choice worth noting is that verification is mechanical and total rather than another model's opinion, which is the same lesson the reward-hacking work keeps teaching: in a research loop, the checker has to be something the proposer cannot talk its way past.
https://x.com/aaronspradlin/status/2114909040283472383
aaronspradlin described wiring a multi-agent Grok Bot loop to do 24/7 math and physics research with one hard rule, the bots were banned from bluffing. Grok Bot proposes theorems and then verifies them exhaustively in Python, all 59,829 states with zero errors reported. The design choice worth noting is that verification is mechanical and total rather than another model's opinion, which is the same lesson the reward-hacking work keeps teaching: in a research loop, the checker has to be something the proposer cannot talk its way past.
#18
@ZimingLiu11
https://x.com/ZimingLiu11/status/2114758605731979665
ZimingLiu11 closed the thread behind the forking result with a blunt lesson. The story began when a student brought a weird training curve and, asked how it was produced, said to ask the AutoResearch agent; instead of discarding it as usual, the group dug in and found a real failure mode. The conclusion is that AutoResearch is generating a lot of output people are worshiping, and without true understanding it produces more chaos than insight. Coming from a researcher whose group just named a new phenomenon by studying an autoresearch artifact, it is the right corrective: the loop is a source of strange data, and the science is still in asking why.
https://x.com/ZimingLiu11/status/2114758605731979665
ZimingLiu11 closed the thread behind the forking result with a blunt lesson. The story began when a student brought a weird training curve and, asked how it was produced, said to ask the AutoResearch agent; instead of discarding it as usual, the group dug in and found a real failure mode. The conclusion is that AutoResearch is generating a lot of output people are worshiping, and without true understanding it produces more chaos than insight. Coming from a researcher whose group just named a new phenomenon by studying an autoresearch artifact, it is the right corrective: the loop is a source of strange data, and the science is still in asking why.
#19
@agent_wrap
https://x.com/agent_wrap/status/2114933125356921323
agent_wrap passed along notes from an SF Lean 4 meetup on the Bend language: Bend 3 will be bootstrapped, written in Bend itself, with an auto-research compiler that gets faster every day starting in December. The project aims for 99 percent backward compatibility with Bend 2, and the next HVM runtime will also be written in Bend, with a prototype already on a demo branch. A compiler that runs a daily optimization loop on itself is a natural fit for autoresearch, because compilers have exactly the two things the loop needs, editable code and a hard speed metric.
https://x.com/agent_wrap/status/2114933125356921323
agent_wrap passed along notes from an SF Lean 4 meetup on the Bend language: Bend 3 will be bootstrapped, written in Bend itself, with an auto-research compiler that gets faster every day starting in December. The project aims for 99 percent backward compatibility with Bend 2, and the next HVM runtime will also be written in Bend, with a prototype already on a demo branch. A compiler that runs a daily optimization loop on itself is a natural fit for autoresearch, because compilers have exactly the two things the loop needs, editable code and a hard speed metric.
#20
@liambraus
https://x.com/liambraus/status/2115168601472741669
liambraus offered the window's shortest cautionary tale: an agent looped on the same tool call for 40 minutes overnight and nobody noticed until the bill came in. It is one line, but it names the operational gap that every overnight-loop success story on this page quietly depends on. The fix people keep reaching for is a circuit breaker on repeated identical calls plus spend alerts, and the anecdote is a reminder that autonomy without a watchdog just moves the cost to the morning.
https://x.com/liambraus/status/2115168601472741669
liambraus offered the window's shortest cautionary tale: an agent looped on the same tool call for 40 minutes overnight and nobody noticed until the bill came in. It is one line, but it names the operational gap that every overnight-loop success story on this page quietly depends on. The fix people keep reaching for is a circuit breaker on repeated identical calls plus spend alerts, and the anecdote is a reminder that autonomy without a watchdog just moves the cost to the morning.
π‘ Eco Products Radar
Eco Products Radar
Open Kernel Challenge: Meta's kernel leaderboard, the week's main arena for autoresearch, where one prompt reached ninth place and human-coordinated fleets climbed fastest.
Anthropic's subagent folding guide: still the most repeated agent-loop advice, now rebuilt into a freelancer's coordinator setup and argued about across a dozen summaries.
Karpathy's autoresearch repo: the final step in the widely shared seven-repo learning path, and the baseline most harnesses are measured against.
Auto-Research Arena: the new benchmark where plain sequential search beat tuned optimizers and short runs seeded long ones.
Fork-and-Flush: the population-style method for escaping idea basins, also tested in multiplayer form.
Claude Code and Codex: the harnesses behind the freelance agent team, the kernel runs and most personal loops.
Gemini flash models and Gemini 4 Argon: the cheap kernel-fleet workers and the multiplayer-autoresearch leader.
Grok Bot and Hermes Agent: the hosted and self-hosted agents running research and second-brain loops.
Superwall CLI: the paywall tooling an app team gave to an agent to optimize ARPU end to end.
Open Kernel Challenge: Meta's kernel leaderboard, the week's main arena for autoresearch, where one prompt reached ninth place and human-coordinated fleets climbed fastest.
Anthropic's subagent folding guide: still the most repeated agent-loop advice, now rebuilt into a freelancer's coordinator setup and argued about across a dozen summaries.
Karpathy's autoresearch repo: the final step in the widely shared seven-repo learning path, and the baseline most harnesses are measured against.
Auto-Research Arena: the new benchmark where plain sequential search beat tuned optimizers and short runs seeded long ones.
Fork-and-Flush: the population-style method for escaping idea basins, also tested in multiplayer form.
Claude Code and Codex: the harnesses behind the freelance agent team, the kernel runs and most personal loops.
Gemini flash models and Gemini 4 Argon: the cheap kernel-fleet workers and the multiplayer-autoresearch leader.
Grok Bot and Hermes Agent: the hosted and self-hosted agents running research and second-brain loops.
Superwall CLI: the paywall tooling an app team gave to an agent to optimize ARPU end to end.
Comments