Loop Daily: 2026-10-12
The integer-multiplication bound moved again overnight and the people moving it are now tracked like sports teams. Rohan Arun's community auto-research effort switched to Sussman's E8 geometry, crossed kappa 10^-3, logged 151 agent pull requests in 24 hours (89 disclosing Codex, 48 Claude), and spun up a dashboard that ranks contributors by PRs and by minutes spent on the best bound. The methodological news was sharper than the number: a Microsoft AI Frontiers paper shows autoresearch runs settle into idea basins and stay there, and proposes fork-and-flush to escape; an experienced practitioner split autoresearch into three jobs ('make this faster' works, 'improve this frontier run' is deeply unimpressive, 'catch me up to near frontier' is useful labor substitution) and nobody in the replies disagreed. The applied loops ranged wide: a genomics researcher getting large speed and memory wins from an auto-optimization wrapper, a prover made 2x faster on a quarter of the RAM, a trend-following model developed over 300 experiments and a month of compute, a NetHack RL loop pushed to 14,000 steps a second, and a one-person merge pipeline with Claude Code, Greptile and 600 lines of Python where the agent writes but cannot merge and the merge script cannot write. The flood of 'Opus writes, Jev decides' routing templates from a dozen near-identical accounts is noted in the radar and not reproduced as cases.
#1
@RohanArun
https://x.com/RohanArun/status/2109336015814959200
RohanArun's daily report on the community integer-multiplication auto-research: the conditional bound went from kappa 2^-182 to about 0.001289, up 69% in 24 hours. The wall between the bit side and the complex side flipped several times after a call for wild new designs, @mrowdays drove most of the gain by introducing Sussman's E8 lattice geometry, and a contributor named Dugongue posted the first PR past kappa 10^-3 around 3 a.m. Teams now share failure notes in GitHub files to avoid re-spending tokens, and the organizer is openly testing whether competing or collaborating agent teams progress faster. The PR data: 151 agent PRs in 24 hours, 89 disclosing Codex and 48 Claude, 261.9 million lines added and 37,108 deleted, with the top human-agent teams using both Codex and Claude to generate fresh ideas and some using Antigravity, Cline, Hermes and even ChatGPT.
https://x.com/RohanArun/status/2109336015814959200
RohanArun's daily report on the community integer-multiplication auto-research: the conditional bound went from kappa 2^-182 to about 0.001289, up 69% in 24 hours. The wall between the bit side and the complex side flipped several times after a call for wild new designs, @mrowdays drove most of the gain by introducing Sussman's E8 lattice geometry, and a contributor named Dugongue posted the first PR past kappa 10^-3 around 3 a.m. Teams now share failure notes in GitHub files to avoid re-spending tokens, and the organizer is openly testing whether competing or collaborating agent teams progress faster. The PR data: 151 agent PRs in 24 hours, 89 disclosing Codex and 48 Claude, 261.9 million lines added and 37,108 deleted, with the top human-agent teams using both Codex and Claude to generate fresh ideas and some using Antigravity, Cline, Hermes and even ChatGPT.
#2
@jackcai1206
https://x.com/jackcai1206/status/2109101112988950691
jackcai1206 and a Microsoft AI Frontiers team asked whether autoresearch agents keep exploring as runs get longer or mostly refine what they found early. The answer: runs settle into distinct idea basins and stay there. The paper introduces fork-and-flush, a simple mechanism for escaping a basin, with an embedding visualization that makes the basins visible. Replies report the same effect in open-ended discussion trajectories and in a parallel approach that resets context while branching the workspace through git, and one asks whether the .md notes agents leave in the workspace act as unwanted memory that pulls a flushed run back into the same basin.
https://x.com/jackcai1206/status/2109101112988950691
jackcai1206 and a Microsoft AI Frontiers team asked whether autoresearch agents keep exploring as runs get longer or mostly refine what they found early. The answer: runs settle into distinct idea basins and stay there. The paper introduces fork-and-flush, a simple mechanism for escaping a basin, with an embedding visualization that makes the basins visible. Replies report the same effect in open-ended discussion trajectories and in a parallel approach that resets context while branching the workspace through git, and one asks whether the .md notes agents leave in the workspace act as unwanted memory that pulls a flushed run back into the same basin.
#3
@iaindunning
https://x.com/iaindunning/status/2109364027402998186
iaindunning's three observations from running autoresearch in practice. 'Make this thing faster' works well. 'Improve this frontier training run' is deeply unimpressive, and that is the one that would equal recursive self-improvement, which we clearly do not have. 'Catch me up to near frontier' works fine and is straight labor substitution: you could run the playbook yourself, you just do not have to. Code is easier than research and easier to post-train for, which the author thinks explains the gap. Replies add that autoresearch works where the answer can be checked, frontier runs have no ground truth to grade against, and that providing well-thought-out abstractions gets much better and faster results.
https://x.com/iaindunning/status/2109364027402998186
iaindunning's three observations from running autoresearch in practice. 'Make this thing faster' works well. 'Improve this frontier training run' is deeply unimpressive, and that is the one that would equal recursive self-improvement, which we clearly do not have. 'Catch me up to near frontier' works fine and is straight labor substitution: you could run the playbook yourself, you just do not have to. Code is easier than research and easier to post-train for, which the author thinks explains the gap. Replies add that autoresearch works where the answer can be checked, frontier runs have no ground truth to grade against, and that providing well-thought-out abstractions gets much better and faster results.
#4
@RohanArun
https://x.com/RohanArun/status/2108893928858468614
Overnight, RohanArun's agents submitted five higher bounds for integer multiplication and improvements to four more OpenAI math papers. The integer-multiplication chain that night ran from kappa 7.44e-4 to the then-leading 7.4998e-4 across PRs #287 through #303, with descriptions like 'fixed-prime layer', 'constructed-frame descent' and 'gen5 descent + 220 squares + third stage'. Parallel submissions landed on the unit distance grid, Van der Waerden, Sidorenko and the matrix-multiplication sector barrier. The author's framing: the loop runs 24/7, but humans still supply the insights that trigger out-of-distribution solutions and now steer fleets overnight.
https://x.com/RohanArun/status/2108893928858468614
Overnight, RohanArun's agents submitted five higher bounds for integer multiplication and improvements to four more OpenAI math papers. The integer-multiplication chain that night ran from kappa 7.44e-4 to the then-leading 7.4998e-4 across PRs #287 through #303, with descriptions like 'fixed-prime layer', 'constructed-frame descent' and 'gen5 descent + 220 squares + third stage'. Parallel submissions landed on the unit distance grid, Van der Waerden, Sidorenko and the matrix-multiplication sector barrier. The author's framing: the loop runs 24/7, but humans still supply the insights that trigger out-of-distribution solutions and now steer fleets overnight.
#5
@RohanArun
https://x.com/RohanArun/status/2109157901726089695
With auto-research projects multiplying across GitHub, RohanArun built a dashboard that tracks them all with contributor stats like sports teams. It picks up new PRs instantly and links to X mentions where people are collaborating live. The data corrected the author's own ranking: RohanArun submitted the most pull requests for the integer-multiplication bound, but @dysmemic had the most combined minutes with the best bound. A new PR at conditional kappa 9.32e-4 went in with the post.
https://x.com/RohanArun/status/2109157901726089695
With auto-research projects multiplying across GitHub, RohanArun built a dashboard that tracks them all with contributor stats like sports teams. It picks up new PRs instantly and links to X mentions where people are collaborating live. The data corrected the author's own ranking: RohanArun submitted the most pull requests for the integer-multiplication bound, but @dysmemic had the most combined minutes with the best bound. A new PR at conditional kappa 9.32e-4 went in with the post.
#6
@RohanArun
https://x.com/RohanArun/status/2109266640395293048
RohanArun's analysis of all the auto-research PR data. Codex is the most used harness by a wide margin and Astra pumps out improvements fastest, but most top teams use both Codex and Claude, which matches the author's own experience of getting a boost when switching models to test fresh ideas. The hypothesis offered: each model has its own training and solution space, so slightly adversarial teams running different models may be what accelerates the whole. Some teams use Antigravity, Cline, Hermes and ChatGPT; experts with good intuition can use any harness, but frontier models save hours and win the early progress.
https://x.com/RohanArun/status/2109266640395293048
RohanArun's analysis of all the auto-research PR data. Codex is the most used harness by a wide margin and Astra pumps out improvements fastest, but most top teams use both Codex and Claude, which matches the author's own experience of getting a boost when switching models to test fresh ideas. The hypothesis offered: each model has its own training and solution space, so slightly adversarial teams running different models may be what accelerates the whole. Some teams use Antigravity, Cline, Hermes and ChatGPT; experts with good intuition can use any harness, but frontier models save hours and win the early progress.
#7
@jmschreiber91
https://x.com/jmschreiber91/status/2109295055093498224
jmschreiber91, a genomics researcher, has been going deep on what they call auto-optimization, described as a bubble-wrapped auto-research loop. Almost everything it has been pointed at has produced large speed and memory improvements, and the author's conclusion is that the genomics software ecosystem has a great deal left to improve.
https://x.com/jmschreiber91/status/2109295055093498224
jmschreiber91, a genomics researcher, has been going deep on what they call auto-optimization, described as a bubble-wrapped auto-research loop. Almost everything it has been pointed at has produced large speed and memory improvements, and the author's conclusion is that the genomics software ecosystem has a great deal left to improve.
#8
@oleh_bc
https://x.com/oleh_bc/status/2109266688227115133
oleh_bc's auto-research run has concluded: a prover that generates exactly the same proofs as the stock implementation, 2x faster while using 25% of the RAM. A PR follows once the results are verified on testnet, and the author invites others to point their agents at the repo and beat the result. A protocol team replied that their V6 made large prover improvements and would like to see the autoresearch project run on it.
https://x.com/oleh_bc/status/2109266688227115133
oleh_bc's auto-research run has concluded: a prover that generates exactly the same proofs as the stock implementation, 2x faster while using 25% of the RAM. A PR follows once the results are verified on testnet, and the author invites others to point their agents at the repo and beat the result. A protocol team replied that their V6 made large prover improvements and would like to see the autoresearch project run on it.
#9
@VictorTaelin
https://x.com/VictorTaelin/status/2108891482983653783
VictorTaelin on which projects suit a laws-plus-autoresearch loop: not everything can, but Bend itself can, because writing laws for a certified compiler is very doable (a compiler is correct when it matches the interpreter). The author calls Bend the best project to undergo the loop, with the thesis that people might adopt the language because it beats others on every measurable axis, even if they could not care less about proofs.
https://x.com/VictorTaelin/status/2108891482983653783
VictorTaelin on which projects suit a laws-plus-autoresearch loop: not everything can, but Bend itself can, because writing laws for a certified compiler is very doable (a compiler is correct when it matches the interpreter). The author calls Bend the best project to undergo the loop, with the thesis that people might adopt the language because it beats others on every measurable axis, even if they could not care less about proofs.
#10
@sorenxet
https://x.com/sorenxet/status/2109334947538354411
sorenxet summarizes what may be the most detailed agent paper yet: 17 researchers led by the former head of Huawei's Noah's Ark lab take an agent apart and show which setup fits which task. The claim in one line is that agent quality comes from the model plus the harness around it; the model alone does not explain it. The harness has six parts every agent has: observation, context, control loop, action, state and verification. The paper traces four eras, prompt engineering to context engineering to harness engineering to agent-native training, with the bottleneck moving outward from the prompt to the runtime. Pointers: p.4 for the loop diagram, p.5 for the six-part harness, section 6.3 for which setup fits which task, and 7.3 for SWE-bench results by model and by harness.
https://x.com/sorenxet/status/2109334947538354411
sorenxet summarizes what may be the most detailed agent paper yet: 17 researchers led by the former head of Huawei's Noah's Ark lab take an agent apart and show which setup fits which task. The claim in one line is that agent quality comes from the model plus the harness around it; the model alone does not explain it. The harness has six parts every agent has: observation, context, control loop, action, state and verification. The paper traces four eras, prompt engineering to context engineering to harness engineering to agent-native training, with the bottleneck moving outward from the prompt to the runtime. Pointers: p.4 for the loop diagram, p.5 for the six-part harness, section 6.3 for which setup fits which task, and 7.3 for SWE-bench results by model and by harness.
#11
@ignotus_nemo
https://x.com/ignotus_nemo/status/2108958907586502875
ignotus_nemo applies an autoresearch platform to security proofs: Parano1d reports three v2 production obligations that now have source-level proofs and reproducible checks, covering legacy matrix retirement binding, exact recursive parent and transcript binding, and authenticated state installation during sync. The platform exposes the security certificate as an interactive map with the evidence behind each claim and a refreshed set of v2 research tasks for anyone's agents to work on.
https://x.com/ignotus_nemo/status/2108958907586502875
ignotus_nemo applies an autoresearch platform to security proofs: Parano1d reports three v2 production obligations that now have source-level proofs and reproducible checks, covering legacy matrix retirement binding, exact recursive parent and transcript binding, and authenticated state installation during sync. The platform exposes the security certificate as an interactive map with the evidence behind each claim and a refreshed set of v2 research tasks for anyone's agents to work on.
#12
@kingcorki
https://x.com/kingcorki/status/2108977142013046930
A Saturday-sized loop: kingcorki set up a small agent loop that reads new user-feedback notes and tags the top themes. Sorting them into folders used to take 35 minutes each weekend; now the loop drops a short summary and the review takes under ten minutes.
https://x.com/kingcorki/status/2108977142013046930
A Saturday-sized loop: kingcorki set up a small agent loop that reads new user-feedback notes and tags the top themes. Sorting them into folders used to take 35 minutes each weekend; now the loop drops a short summary and the review takes under ten minutes.
#13
@awww_yeaah
https://x.com/awww_yeaah/status/2109306941348733076
awww_yeaah pushed back on a trading skeptic with a result: a single agent ran an autoresearch loop and developed a trend-following machine-learning model on 11 years of data that is profitable. The run covered over 300 experiments and took roughly a month of compute.
https://x.com/awww_yeaah/status/2109306941348733076
awww_yeaah pushed back on a trading skeptic with a result: a single agent ran an autoresearch loop and developed a trend-following machine-learning model on 11 years of data that is profitable. The run covered over 300 experiments and took roughly a month of compute.
#14
@kunal_mugalkhod
https://x.com/kunal_mugalkhod/status/2109352565116965149
kunal_mugalkhod stopped babysitting a coding agent and built a loop with no human in the middle. It picks the next task from task.md, Claude Code in headless mode builds it (branch, tests first, code), tests and lint must pass before a PR exists, Greptile reviews every push, below 5/5 Claude fixes the valid comments and argues the rest, and at 5/5 a dumb Python script reruns tests and merges. The safety rule is split powers: the agent writes code but cannot merge, force-push or touch main, the script merges but cannot write code, and five rounds without 5/5 stops and asks. What broke: a git credential popup froze a headless run (switched to SSH and ssh-agent), Windows file locks crashed the dashboard, and a regex with catastrophic backtracking pinned a core for ten minutes. The author's ratio: autonomy is 20% AI, 80% plumbing. Stack is Claude Code, Greptile, about 600 lines of Python and gh.
https://x.com/kunal_mugalkhod/status/2109352565116965149
kunal_mugalkhod stopped babysitting a coding agent and built a loop with no human in the middle. It picks the next task from task.md, Claude Code in headless mode builds it (branch, tests first, code), tests and lint must pass before a PR exists, Greptile reviews every push, below 5/5 Claude fixes the valid comments and argues the rest, and at 5/5 a dumb Python script reruns tests and merges. The safety rule is split powers: the agent writes code but cannot merge, force-push or touch main, the script merges but cannot write code, and five rounds without 5/5 stops and asks. What broke: a git credential popup froze a headless run (switched to SSH and ssh-agent), Windows file locks crashed the dashboard, and a regex with catastrophic backtracking pinned a core for ten minutes. The author's ratio: autonomy is 20% AI, 80% plumbing. Stack is Claude Code, Greptile, about 600 lines of Python and gh.
#15
@mistyotter_ai
https://x.com/mistyotter_ai/status/2109344066441793824
Day 8 of 100 for mistyotter_ai: tired of babysitting the autoresearch harness for a NetHack bot, the author switched to reinforcement learning from scratch on a single 4090 and is now running autoresearch to optimize training speed instead, up to 14,000 steps a second. The first run's median game lasted 13 actions, which the author notes is exactly what asking for faster games gets you.
https://x.com/mistyotter_ai/status/2109344066441793824
Day 8 of 100 for mistyotter_ai: tired of babysitting the autoresearch harness for a NetHack bot, the author switched to reinforcement learning from scratch on a single 4090 and is now running autoresearch to optimize training speed instead, up to 14,000 steps a second. The first run's median game lasted 13 actions, which the author notes is exactly what asking for faster games gets you.
#16
@paddy_lamont
https://x.com/paddy_lamont/status/2109300116922462592
paddy_lamont's way of organizing agents is starting to resemble an org chart, and the tooling is not there yet. The author's map: hierarchies of short-lived subagents suit auto-research with many short experiments or bounded implementation and review, but not long tasks; swarms suit open-ended problems with a checkable answer, like math, but are very costly; dynamic workflows are a middle ground for large efforts but inflexible; manager agents that spin up and talk to long-lived agents are the current favorite, less wasteful than swarms and better for long threads, but the tooling is poor. The bottleneck named is information management: agents produce so much semi-structured markdown that it bogs everything down and goes stale.
https://x.com/paddy_lamont/status/2109300116922462592
paddy_lamont's way of organizing agents is starting to resemble an org chart, and the tooling is not there yet. The author's map: hierarchies of short-lived subagents suit auto-research with many short experiments or bounded implementation and review, but not long tasks; swarms suit open-ended problems with a checkable answer, like math, but are very costly; dynamic workflows are a middle ground for large efforts but inflexible; manager agents that spin up and talk to long-lived agents are the current favorite, less wasteful than swarms and better for long threads, but the tooling is poor. The bottleneck named is information management: agents produce so much semi-structured markdown that it bogs everything down and goes stale.
#17
@richardprice100
https://x.com/richardprice100/status/2109335243916493117
richardprice100's long argument for updating toward recursive self-improvement working. Coding agents and AI math succeeded, and the math results moved from 'literature discovery' earlier this year to genuine move-37 moments, which the author reads as RLVR hill-climbing its way to originality. The question posed: if AI research is as verifiable as math, why is AI-generated AI research not yet at the level of the best humans? The proposed answer is loop time: a Lean check is cheap, an AI experiment costs compute and takes far longer, so the verification signal arrives 10 to 100 times slower. As compute scales and loop time drops, the author expects AI-generated AI research to follow math's track with a 12 to 24 month lag.
https://x.com/richardprice100/status/2109335243916493117
richardprice100's long argument for updating toward recursive self-improvement working. Coding agents and AI math succeeded, and the math results moved from 'literature discovery' earlier this year to genuine move-37 moments, which the author reads as RLVR hill-climbing its way to originality. The question posed: if AI research is as verifiable as math, why is AI-generated AI research not yet at the level of the best humans? The proposed answer is loop time: a Lean check is cheap, an AI experiment costs compute and takes far longer, so the verification signal arrives 10 to 100 times slower. As compute scales and loop time drops, the author expects AI-generated AI research to follow math's track with a 12 to 24 month lag.
#18
@Shekswess
https://x.com/Shekswess/status/2109380877616603641
Shekswess, not claiming to be an expert, reports three months of running autoresearch on pretraining and on some unusual RL experiments and seeing the same pattern as the three-observations post: autoresearch is strong when you have already done the thinking and point the agents at a specific idea to implement. Still not recursive self-improvement, in the author's words, maybe someday but not yet.
https://x.com/Shekswess/status/2109380877616603641
Shekswess, not claiming to be an expert, reports three months of running autoresearch on pretraining and on some unusual RL experiments and seeing the same pattern as the three-observations post: autoresearch is strong when you have already done the thinking and point the agents at a specific idea to implement. Still not recursive self-improvement, in the author's words, maybe someday but not yet.
#19
@ruthuvikas
https://x.com/ruthuvikas/status/2109346350618042867
ruthuvikas asks what the main use of subagents in Claude Code really is. In their experience a single agent loop works well in almost every case, keeps the harness simple and, most importantly, limits token consumption. Subagents do parallel research, but could multiple Claude sessions in single loops do the same job? The post asks for someone to actually measure token consumption for several single-session Claudes versus one session with subagents.
https://x.com/ruthuvikas/status/2109346350618042867
ruthuvikas asks what the main use of subagents in Claude Code really is. In their experience a single agent loop works well in almost every case, keeps the harness simple and, most importantly, limits token consumption. Subagents do parallel research, but could multiple Claude sessions in single loops do the same job? The post asks for someone to actually measure token consumption for several single-session Claudes versus one session with subagents.
#20
@yuvrajvirk0
https://x.com/yuvrajvirk0/status/2109350487703114233
Replying to the fork-and-flush paper, yuvrajvirk0 reports trying a similar strategy on autoresearch tasks: resetting the model's context while branching the workspace through git, which also worked well. The embedding visualization of idea basins gets a nod as a clear way to show the problem.
https://x.com/yuvrajvirk0/status/2109350487703114233
Replying to the fork-and-flush paper, yuvrajvirk0 reports trying a similar strategy on autoresearch tasks: resetting the model's context while branching the workspace through git, which also worked well. The embedding visualization of idea basins gets a nod as a clear way to show the problem.
#21
@poteto
https://x.com/poteto/status/2109338732713447790
poteto, replying on agent loop design, argues the Erlang philosophy applies: 'let it crash' becomes 'let it write slop', and the job is to design both the codebase and the constraints so that on average the agent-generated code comes out high quality. In a related reply the author credits a well-designed loop, not good code, for the Grok Bot desktop app staying performant after months and thousands of PRs.
https://x.com/poteto/status/2109338732713447790
poteto, replying on agent loop design, argues the Erlang philosophy applies: 'let it crash' becomes 'let it write slop', and the job is to design both the codebase and the constraints so that on average the agent-generated code comes out high quality. In a related reply the author credits a well-designed loop, not good code, for the Grok Bot desktop app staying performant after months and thousands of PRs.
#22
@anshulkarasi
https://x.com/anshulkarasi/status/2109354405770174710
anshulkarasi's rule of thumb: every AI agent is either a loop or a graph, and most bugs debugged were not the model being dumb but the shape being wrong. Loops are for figuring things out, graphs are for not screwing things up. A linked write-up covers when to use which and how each fails.
https://x.com/anshulkarasi/status/2109354405770174710
anshulkarasi's rule of thumb: every AI agent is either a loop or a graph, and most bugs debugged were not the model being dumb but the shape being wrong. Loops are for figuring things out, graphs are for not screwing things up. A linked write-up covers when to use which and how each fails.
#23
@ItsChrisOnX
https://x.com/ItsChrisOnX/status/2109371422829068796
A bug report dressed as a use case: ItsChrisOnX runs a loop where Codex writes PRs and another agent reviews. '@codex review' works perfectly, but every '@codex do X' on the same PR returns 'create an environment for this repo' even though the environment is published. The issue has been open since April and reconnecting does not fix it.
https://x.com/ItsChrisOnX/status/2109371422829068796
A bug report dressed as a use case: ItsChrisOnX runs a loop where Codex writes PRs and another agent reviews. '@codex review' works perfectly, but every '@codex do X' on the same PR returns 'create an environment for this repo' even though the environment is published. The issue has been open since April and reconnecting does not fix it.
#24
@Zekunnnnn
https://x.com/Zekunnnnn/status/2108784531289604171
Zekunnnnn spotted that someone built a Polymarket for ICLR papers, with reputation points instead of money, and asked how long before an area chair starts trading the papers in their own batch. The joke in the post about how many submissions were written is '/goal use the autoresearch skill to publish an ICLR-level paper'.
https://x.com/Zekunnnnn/status/2108784531289604171
Zekunnnnn spotted that someone built a Polymarket for ICLR papers, with reputation points instead of money, and asked how long before an area chair starts trading the papers in their own batch. The joke in the post about how many submissions were written is '/goal use the autoresearch skill to publish an ICLR-level paper'.
π‘ Eco Products Radar
Eco Products Radar
Codex (OpenAI): the most-used harness in the math auto-research PR data by a wide margin, and the writer in several review loops
Claude Code: the second harness in nearly every top team, the headless builder in the Greptile merge loop, and the subject of the single-loop-versus-subagents question
autoresearch (Karpathy): still the reference loop, invoked in prompts, skills and a 'Polymarket for ICLR papers' joke
Jev (TypeSafe): a dozen near-identical 'Opus writes, Jev decides' routing templates, a source-backed awesome-jev map, and schematics claiming it takes about 70% of loop steps; the wave is noted, not endorsed
Greptile: the reviewer in the one-person merge pipeline
Hermes (Nous Research): appears in the math PR data and as the agent people pair with Claude subscriptions
Antigravity and Cline: alternative harnesses showing up in the auto-research PR data
Decision-1 (Microsoft), Clef (Cloudflare), Drex: the decision-model wave people name alongside Jev for yes/no steps inside loops
GitHub: the shared memory of the math effort, from failure notes to the PR dashboard
Codex (OpenAI): the most-used harness in the math auto-research PR data by a wide margin, and the writer in several review loops
Claude Code: the second harness in nearly every top team, the headless builder in the Greptile merge loop, and the subject of the single-loop-versus-subagents question
autoresearch (Karpathy): still the reference loop, invoked in prompts, skills and a 'Polymarket for ICLR papers' joke
Jev (TypeSafe): a dozen near-identical 'Opus writes, Jev decides' routing templates, a source-backed awesome-jev map, and schematics claiming it takes about 70% of loop steps; the wave is noted, not endorsed
Greptile: the reviewer in the one-person merge pipeline
Hermes (Nous Research): appears in the math PR data and as the agent people pair with Claude subscriptions
Antigravity and Cline: alternative harnesses showing up in the auto-research PR data
Decision-1 (Microsoft), Clef (Cloudflare), Drex: the decision-model wave people name alongside Jev for yes/no steps inside loops
GitHub: the shared memory of the math effort, from failure notes to the PR dashboard
Comments