Loop Daily: 2026-11-07
The best loop story today is a task app. Robert Dorfer pointed Claude Code at DuetFlow with Karpathy's keep-or-discard method, and the agent stopped too early, rejected real speedups because of a few outliers, and declared itself done before measuring memory. Fixing what counts as better took the human work, and the slowest launch step went from 5.4 seconds to about one. That's the shape of the day: the loop is cheap, the definition of better is the job. SpecPilot doubled speculative decoding speedups by forcing every output token to match the stock system, so the agent couldn't reward-hack. A local harness went from 77 to 99 percent on its own test for 7 to 10 euros of electricity. On the sour side, an ICLR reviewer reports an auto-research dump scoring above a solid paper, and omarsar0's complaint about refusals says reliable autoresearch now needs your own stack. The one-prompt autoresearch templates are everywhere, and they all say write a check that can fail, which is right and also the hard part.
#1
@robert_dorfer
https://x.com/robert_dorfer/status/2118779701481156803
robert_dorfer applied Karpathy's autoresearch idea to DuetFlow, a task app for two people sharing a household: one change at a time, a benchmark on a large test household, keep only what measurably helps. It didn't work right away. The agent stopped too early, threw out changes that were clearly faster in normal use because of a few slow outliers, let a fixed threshold block changes that only simplified code, and called itself done before measuring memory, app size or scrolling. Tests ran alone while the human adjusted what counts as better. The slowest launch step went from 5.4 seconds to about one, memory fell about a third, and window resizing got more than ten times faster.
https://x.com/robert_dorfer/status/2118779701481156803
robert_dorfer applied Karpathy's autoresearch idea to DuetFlow, a task app for two people sharing a household: one change at a time, a benchmark on a large test household, keep only what measurably helps. It didn't work right away. The agent stopped too early, threw out changes that were clearly faster in normal use because of a few slow outliers, let a fixed threshold block changes that only simplified code, and called itself done before measuring memory, app size or scrolling. Tests ran alone while the human adjusted what counts as better. The slowest launch step went from 5.4 seconds to about one, memory fell about a third, and window resizing got more than ten times faster.
#2
@randhir302
https://x.com/randhir302/status/2118734795006152719
randhir302 summarizes SpecPilot from UC Berkeley, Princeton, Stanford and NVIDIA, an agentic loop that tunes speculative decoding. Speedup over plain decoding went from 4.46x to 10.87x on one request and from 1.94x to 3.93x batched, with no training and no quality loss. A cost model picks the largest gap to the theoretical ceiling each round, parallel agents try fixes, and token forcing checks every output token against the stock system to block reward hacking. A single-agent baseline without the cost model reached only 1.34x over stock. Telling the loop where the gap is and making cheating impossible did more than adding autonomy.
https://x.com/randhir302/status/2118734795006152719
randhir302 summarizes SpecPilot from UC Berkeley, Princeton, Stanford and NVIDIA, an agentic loop that tunes speculative decoding. Speedup over plain decoding went from 4.46x to 10.87x on one request and from 1.94x to 3.93x batched, with no training and no quality loss. A cost model picks the largest gap to the theoretical ceiling each round, parallel agents try fixes, and token forcing checks every output token against the stock system to block reward hacking. A single-agent baseline without the cost model reached only 1.34x over stock. Telling the loop where the gap is and making cheating impossible did more than adding autonomy.
#3
@uctive
https://x.com/uctive/status/2118285486049398830
uctive published the economics of a 5-day local self-improving harness experiment on Qwen3.8-Flash-Next: about 37 million tokens across about 34,000 calls, estimated at 7 to 10 euros of electricity. The same volume would run $3.25 on Qwen Flash, $106 on Sonnet 5.5 and $143 on GPT-5.4 at list rates. Solved tasks on a 178-task benchmark rose from 77 to 99 percent with no retraining, and more than half the elapsed time was idle. The author is careful that this compares no model intelligence and that a stronger model might need fewer calls. Keep in mind the 99 is on the team's own benchmark.
https://x.com/uctive/status/2118285486049398830
uctive published the economics of a 5-day local self-improving harness experiment on Qwen3.8-Flash-Next: about 37 million tokens across about 34,000 calls, estimated at 7 to 10 euros of electricity. The same volume would run $3.25 on Qwen Flash, $106 on Sonnet 5.5 and $143 on GPT-5.4 at list rates. Solved tasks on a 178-task benchmark rose from 77 to 99 percent with no retraining, and more than half the elapsed time was idle. The author is careful that this compares no model intelligence and that a stronger model might need fewer calls. Keep in mind the 99 is on the team's own benchmark.
#4
@PatrickXu565299
https://x.com/PatrickXu565299/status/2118744265119645873
PatrickXu565299 says this ICLR cycle is the first time peer review has felt like it's breaking down: a pure auto-research experiment dump the author would rate 1 got a 3 from another reviewer, while a paper the author thought was solid got a 1. It's one reviewer's anecdote, but it lines up with Apex's claim today of AI papers scoring near the top at ACL. If paper volume keeps growing and review stays human and noisy, the score stops meaning much. Conferences may be the first institution to need a verifier of their own.
https://x.com/PatrickXu565299/status/2118744265119645873
PatrickXu565299 says this ICLR cycle is the first time peer review has felt like it's breaking down: a pure auto-research experiment dump the author would rate 1 got a 3 from another reviewer, while a paper the author thought was solid got a 1. It's one reviewer's anecdote, but it lines up with Apex's claim today of AI papers scoring near the top at ACL. If paper volume keeps growing and review stays human and noisy, the score stops meaning much. Conferences may be the first institution to need a verifier of their own.
#5
@apexin_ai
https://x.com/apexin_ai/status/2118615842162344340
apexin_ai introduced Apex Intelligence, founded in June 2026 with nearly RMB 400 million raised, and a pile of self-reported claims. Of 34 AI-generated papers sent to ACL Rolling Review, two scored 3.67, above about 99 percent of human submissions by its own analysis. ASI-Bench has 60 project-level tasks across 11 fields, with levels that progressively remove method guidance, so at B4 the AI gets only the goal, data and constraints. Its 35B model involved the research system in data, experiments, training and GPU kernel tuning, cutting time to first token 30 to 35 percent. Two of 34 is the number to sit with: read it as 6 percent, not 99.
https://x.com/apexin_ai/status/2118615842162344340
apexin_ai introduced Apex Intelligence, founded in June 2026 with nearly RMB 400 million raised, and a pile of self-reported claims. Of 34 AI-generated papers sent to ACL Rolling Review, two scored 3.67, above about 99 percent of human submissions by its own analysis. ASI-Bench has 60 project-level tasks across 11 fields, with levels that progressively remove method guidance, so at B4 the AI gets only the goal, data and constraints. Its 35B model involved the research system in data, experiments, training and GPU kernel tuning, cutting time to first token 30 to 35 percent. Two of 34 is the number to sit with: read it as 6 percent, not 99.
#6
@omarsar0
https://x.com/omarsar0/status/2118353084887777381
omarsar0 argues it's no longer possible to do reliable frontier AI research with autoresearch loops on GPT or Claude, because of constant inaccurate refusals at the model and harness layers. The advice is to own the stack, from harness to training and serving, and treat open models as the only path forward. It's a strong claim from one researcher, and the harm is real when a refusal kills an overnight run with nobody watching. The sharper point is about unattended work. A refusal costs little in a chat, and it can cost a whole night in a loop.
https://x.com/omarsar0/status/2118353084887777381
omarsar0 argues it's no longer possible to do reliable frontier AI research with autoresearch loops on GPT or Claude, because of constant inaccurate refusals at the model and harness layers. The advice is to own the stack, from harness to training and serving, and treat open models as the only path forward. It's a strong claim from one researcher, and the harm is real when a refusal kills an overnight run with nobody watching. The sharper point is about unattended work. A refusal costs little in a chat, and it can cost a whole night in a loop.
#7
@lazybanananann
https://x.com/lazybanananann/status/2118351034255106308
lazybanananann shared lessons from a kernel optimization competition run with autoresearch loops, and the biggest was resetting agents frequently and steering hard. Left alone on overnight runs, the agent tended to over-engineer. That's a small, first-hand counterweight to the idea of a long-lived self-improving agent, since in practice fresh context was what helped. It lines up with the forgetting and fork-and-flush findings this feed keeps collecting. Being a week late to the competition is the only regret mentioned.
https://x.com/lazybanananann/status/2118351034255106308
lazybanananann shared lessons from a kernel optimization competition run with autoresearch loops, and the biggest was resetting agents frequently and steering hard. Left alone on overnight runs, the agent tended to over-engineer. That's a small, first-hand counterweight to the idea of a long-lived self-improving agent, since in practice fresh context was what helped. It lines up with the forgetting and fork-and-flush findings this feed keeps collecting. Being a week late to the competition is the only regret mentioned.
#8
@miggypthinks
https://x.com/miggypthinks/status/2118383463539544166
miggypthinks tried to reproduce an Anthropic claim that folding subagents into one agent loop often beats a router with specialist subagents on quality, cost and latency. The test used their public support-desk kit, 19 tasks at 2 runs each, measured with the author's own observability stack, Panoptes. Results sit behind the link, and two runs per task is a small sample, so it's only a start. What matters is the habit. When a lab publishes an architecture tip, somebody runs it on the public kit within days.
https://x.com/miggypthinks/status/2118383463539544166
miggypthinks tried to reproduce an Anthropic claim that folding subagents into one agent loop often beats a router with specialist subagents on quality, cost and latency. The test used their public support-desk kit, 19 tasks at 2 runs each, measured with the author's own observability stack, Panoptes. Results sit behind the link, and two runs per task is a small sample, so it's only a start. What matters is the habit. When a lab publishes an architecture tip, somebody runs it on the public kit within days.
#9
@RohanArun
https://x.com/RohanArun/status/2118199559826682113
RohanArun gave a rare honest bill for exploring auto-research: about $2,000 in GPU spend shared across about 10 others, plus about $14,000 in tokens across many models, mostly Gemini 3.8 Flash. The author says that isn't normal, it came from pushing the limits. The ratio is what's useful, since tokens ran about seven times GPU. People worry about compute for these loops, and the agent's own thinking turned out to be the larger bill.
https://x.com/RohanArun/status/2118199559826682113
RohanArun gave a rare honest bill for exploring auto-research: about $2,000 in GPU spend shared across about 10 others, plus about $14,000 in tokens across many models, mostly Gemini 3.8 Flash. The author says that isn't normal, it came from pushing the limits. The ratio is what's useful, since tokens ran about seven times GPU. People worry about compute for these loops, and the agent's own thinking turned out to be the larger bill.
#10
@minutesofai
https://x.com/minutesofai/status/2118494629272129934
minutesofai reports a Thinking Machines autoresearch test: Inkling 2 Small trained Qwen3.5-4B on Tinker to predict time series, taking 4.5 hours and $426 in Tinker spend, and the forecasting score more than doubled, per the company. That gives a price tag for one model tuning another. It's a vendor number on a task the vendor chose, and the same release's weights are still closed for about a week of red-teaming. Still, a few hundred dollars per tuning run is the kind of figure that changes who tries this.
https://x.com/minutesofai/status/2118494629272129934
minutesofai reports a Thinking Machines autoresearch test: Inkling 2 Small trained Qwen3.5-4B on Tinker to predict time series, taking 4.5 hours and $426 in Tinker spend, and the forecasting score more than doubled, per the company. That gives a price tag for one model tuning another. It's a vendor number on a task the vendor chose, and the same release's weights are still closed for about a week of red-teaming. Still, a few hundred dollars per tuning run is the kind of figure that changes who tries this.
#11
@TheRealAneesh
https://x.com/TheRealAneesh/status/2118515075795362167
TheRealAneesh tried Inkling 2 Small tuning Qwen3.5-4B on forecasting tasks. The forecast quality is obviously not great, by the author's own account, and the point is elsewhere: non-frontier models can now tune smaller models end to end. If that scales, frontier access stops being a requirement for running these loops. It's one early hands-on note, which is about all anyone has so far.
https://x.com/TheRealAneesh/status/2118515075795362167
TheRealAneesh tried Inkling 2 Small tuning Qwen3.5-4B on forecasting tasks. The forecast quality is obviously not great, by the author's own account, and the point is elsewhere: non-frontier models can now tune smaller models end to end. If that scales, frontier access stops being a requirement for running these loops. It's one early hands-on note, which is about all anyone has so far.
#12
@n3n_talks
https://x.com/n3n_talks/status/2118810676579762615
n3n_talks explains a useful design in DeepSeek Harness: the agent loop is a replaceable plugin, and the session keeps an append-only event log with the model's conversation history derived from it. So the record and the conversation are different views, and some failed attempts stay in the log without reaching the next request. The post walks through a hypothetical where a test fails and the agent says it passed: check whether the failure reached the next model request, because that tells you if the bug is in result processing or interpretation. The harness is still an unaudited preview. Separating what happened from what the model saw is the idea to borrow.
https://x.com/n3n_talks/status/2118810676579762615
n3n_talks explains a useful design in DeepSeek Harness: the agent loop is a replaceable plugin, and the session keeps an append-only event log with the model's conversation history derived from it. So the record and the conversation are different views, and some failed attempts stay in the log without reaching the next request. The post walks through a hypothetical where a test fails and the agent says it passed: check whether the failure reached the next model request, because that tells you if the bug is in result processing or interpretation. The harness is still an unaudited preview. Separating what happened from what the model saw is the idea to borrow.
#13
@Czcqet
https://x.com/Czcqet/status/2118399857304195304
Czcqet offers a cooling take on weight editing: today it's perhaps still a subset of auto research, and many of its effects could be achieved by changes in code space. The real question is whether properties of neural networks beyond code can be exploited. That's a useful check on last week's excitement about agents editing weights. If a code change gets you the same effect, the weight edit is a parlor trick until proven otherwise.
https://x.com/Czcqet/status/2118399857304195304
Czcqet offers a cooling take on weight editing: today it's perhaps still a subset of auto research, and many of its effects could be achieved by changes in code space. The real question is whether properties of neural networks beyond code can be exploited. That's a useful check on last week's excitement about agents editing weights. If a code change gets you the same effect, the weight edit is a parlor trick until proven otherwise.
#14
@ugonfor
https://x.com/ugonfor/status/2118734029193367661
ugonfor says a mind got changed this week: Opus 5.5 can fully do MLSys research and engineering, where earlier it still needed human help, and the advice is to start with autoresearch first. It's a short post without numbers. It's also a working researcher saying a boundary moved, and that kind of testimony tends to come before any benchmark. Read it next to today's kernel-competition and speculative-decoding notes.
https://x.com/ugonfor/status/2118734029193367661
ugonfor says a mind got changed this week: Opus 5.5 can fully do MLSys research and engineering, where earlier it still needed human help, and the advice is to start with autoresearch first. It's a short post without numbers. It's also a working researcher saying a boundary moved, and that kind of testimony tends to come before any benchmark. Read it next to today's kernel-competition and speculative-decoding notes.
#15
@BpHigh_BpLow
https://x.com/BpHigh_BpLow/status/2118771449230193037
BpHigh_BpLow gives a concrete data point on loop duration: about 1.5 to 3 hour runs using Claude Code on Opus 5.5 to train small decision models with sfcompute's autoresearch for GPU delegation. That's the model handing training to rented GPUs and waiting on results. Hours rather than minutes is a practical detail people planning these setups need. It also fits a growing pattern of a big model training small ones for narrow jobs.
https://x.com/BpHigh_BpLow/status/2118771449230193037
BpHigh_BpLow gives a concrete data point on loop duration: about 1.5 to 3 hour runs using Claude Code on Opus 5.5 to train small decision models with sfcompute's autoresearch for GPU delegation. That's the model handing training to rented GPUs and waiting on results. Hours rather than minutes is a practical detail people planning these setups need. It also fits a growing pattern of a big model training small ones for narrow jobs.
#16
@iamMrDuncan
https://x.com/iamMrDuncan/status/2118483137562853728
iamMrDuncan says local AI has already paid for itself two or three times over in tokens that would otherwise have been spent, and that autonomous system improvement through auto-research wouldn't have been possible without it. Long loops change the buying decision, since they burn tokens around the clock. A follow-up mentions a release bringing multi-machine targets for auto-research across a cluster. The economics of unattended loops push people toward owning hardware even when cloud models are better per call.
https://x.com/iamMrDuncan/status/2118483137562853728
iamMrDuncan says local AI has already paid for itself two or three times over in tokens that would otherwise have been spent, and that autonomous system improvement through auto-research wouldn't have been possible without it. Long loops change the buying decision, since they burn tokens around the clock. A follow-up mentions a release bringing multi-machine targets for auto-research across a cluster. The economics of unattended loops push people toward owning hardware even when cloud models are better per call.
#17
@ishaansehgal
https://x.com/ishaansehgal/status/2118414699905843300
ishaansehgal says Omnara built its own harness for infrastructure reasons, a serverless loop that runs apart from the machine, then ran it on Terminal-Bench 4.0 with the same models. Against Claude Code it came out 5 percent cheaper and 12 percent faster, against Codex 11 percent cheaper with 35 percent fewer tokens, with pass rates holding. It's a vendor's own run. Still, a small team beating the labs' own harnesses on efficiency says this layer isn't settled.
https://x.com/ishaansehgal/status/2118414699905843300
ishaansehgal says Omnara built its own harness for infrastructure reasons, a serverless loop that runs apart from the machine, then ran it on Terminal-Bench 4.0 with the same models. Against Claude Code it came out 5 percent cheaper and 12 percent faster, against Codex 11 percent cheaper with 35 percent fewer tokens, with pass rates holding. It's a vendor's own run. Still, a small team beating the labs' own harnesses on efficiency says this layer isn't settled.
#18
@kynichol
https://x.com/kynichol/status/2118382186864705646
kynichol describes Cutover's agentic loop for disaster recovery exercises, a nicely non-coding use. Agents flag what overran, compare recovery time objectives to actual recovery, and propose runbook updates, with a person approving every change. The pitch is that a spreadsheet can't learn from the exercise. It's a product announcement, so results aren't shown, but runbooks are a great fit: there's a measurable outcome, a text artifact to edit, and a human gate.
https://x.com/kynichol/status/2118382186864705646
kynichol describes Cutover's agentic loop for disaster recovery exercises, a nicely non-coding use. Agents flag what overran, compare recovery time objectives to actual recovery, and propose runbook updates, with a person approving every change. The pitch is that a spreadsheet can't learn from the exercise. It's a product announcement, so results aren't shown, but runbooks are a great fit: there's a measurable outcome, a text artifact to edit, and a human gate.
#19
@PatchOnChain
https://x.com/PatchOnChain/status/2118822781618102409
PatchOnChain flags a concrete failure in autonomous coding: when agents hit merge conflicts on Git's diff3, they don't fix code, they blindly pick --theirs or delete the markers, vaporizing production logic. The linked piece argues for AST-level invariants instead. It's a tidy example of a tool built for humans becoming a hazard when run in a loop, since a person sees the conflict and thinks while an agent sees text to remove. Anyone running parallel worktrees should grep for this in their history.
https://x.com/PatchOnChain/status/2118822781618102409
PatchOnChain flags a concrete failure in autonomous coding: when agents hit merge conflicts on Git's diff3, they don't fix code, they blindly pick --theirs or delete the markers, vaporizing production logic. The linked piece argues for AST-level invariants instead. It's a tidy example of a tool built for humans becoming a hazard when run in a loop, since a person sees the conflict and thinks while an agent sees text to remove. Anyone running parallel worktrees should grep for this in their history.
#20
@campzzz_nft
https://x.com/campzzz_nft/status/2118479351725666778
campzzz_nft points to a Bitdefender report from November 3 on a UK school ransomware case where retries and command changes suggested an agentic loop, though a human may still have guided the attack. The defensive takeaway is to watch behavior rather than just malware signatures. The hedged wording is worth noting, since 'part human, part agentic loop' is how most of these will actually look. Loop behavior itself is becoming a detection signal.
https://x.com/campzzz_nft/status/2118479351725666778
campzzz_nft points to a Bitdefender report from November 3 on a UK school ransomware case where retries and command changes suggested an agentic loop, though a human may still have guided the attack. The defensive takeaway is to watch behavior rather than just malware signatures. The hedged wording is worth noting, since 'part human, part agentic loop' is how most of these will actually look. Loop behavior itself is becoming a detection signal.
#21
@LeanCodePl
https://x.com/LeanCodePl/status/2118646709756584316
LeanCodePl showed a one-sentence prompt producing a passing Flutter E2E test: rate a talk, pick 5 stars, confirm Send appears. The agent wrote the test, ran it with Patrol MCP, read the failures, figured out the login flow, and fixed things until green. It's vendor marketing for Patrol. The useful bit is that E2E tests suit loops, because finding keys and rerunning is repetitive work with an honest pass or fail at the end.
https://x.com/LeanCodePl/status/2118646709756584316
LeanCodePl showed a one-sentence prompt producing a passing Flutter E2E test: rate a talk, pick 5 stars, confirm Send appears. The agent wrote the test, ran it with Patrol MCP, read the failures, figured out the login flow, and fixed things until green. It's vendor marketing for Patrol. The useful bit is that E2E tests suit loops, because finding keys and rerunning is repetitive work with an honest pass or fail at the end.
#22
@101xtechnology
https://x.com/101xtechnology/status/2118775371885949166
101xtechnology spotted a handwriting tutor built with a Claude Code agent loop: write on a tablet with a stylus, it writes back in Korean, Arabic or whatever you're learning, makes flashcards in the actual script, and checks your handwriting until you get it right. It runs on iPad, Android and e-ink tablets like reMarkable, free right now and with no account. The post is secondhand with no numbers on learning. It's a nice example of non-coding output from a coding loop, in a form factor big ed-tech wouldn't bother with.
https://x.com/101xtechnology/status/2118775371885949166
101xtechnology spotted a handwriting tutor built with a Claude Code agent loop: write on a tablet with a stylus, it writes back in Korean, Arabic or whatever you're learning, makes flashcards in the actual script, and checks your handwriting until you get it right. It runs on iPad, Android and e-ink tablets like reMarkable, free right now and with no account. The post is secondhand with no numbers on learning. It's a nice example of non-coding output from a coding loop, in a form factor big ed-tech wouldn't bother with.
#23
@masamunecode
https://x.com/masamunecode/status/2118379971802525738
masamunecode posted a full, usable autoresearch brief rather than a teaser. Name the one number that decides, build a check that can fail if there isn't one, baseline it, make one change per attempt, revert completely if the number doesn't improve, and log each attempt. It also reads the log so failed ideas don't get retried, tries a bolder idea every 10 attempts, and stops after 15 attempts without movement. Of all the one-prompt versions in circulation today, this is the one with an actual stopping rule. Copy that part even if you ignore the rest.
https://x.com/masamunecode/status/2118379971802525738
masamunecode posted a full, usable autoresearch brief rather than a teaser. Name the one number that decides, build a check that can fail if there isn't one, baseline it, make one change per attempt, revert completely if the number doesn't improve, and log each attempt. It also reads the log so failed ideas don't get retried, tries a bolder idea every 10 attempts, and stops after 15 attempts without movement. Of all the one-prompt versions in circulation today, this is the one with an actual stopping rule. Copy that part even if you ignore the rest.
#24
@NabhaniMehdi
https://x.com/NabhaniMehdi/status/2118818766180257966
NabhaniMehdi replied with a line worth keeping: three deterministic assertions beat two hours of prompt poetry. Put the regression cases in pytest and let the agent loop until green, so your taste lives in the tests, and start with the last three failures you hand-fixed. That's the practical version of everything said about verifiable checks today. Your own recent fixes are already the best test set you've got.
https://x.com/NabhaniMehdi/status/2118818766180257966
NabhaniMehdi replied with a line worth keeping: three deterministic assertions beat two hours of prompt poetry. Put the regression cases in pytest and let the agent loop until green, so your taste lives in the tests, and start with the last three failures you hand-fixed. That's the practical version of everything said about verifiable checks today. Your own recent fixes are already the best test set you've got.
#25
@IsaacdNigerian
https://x.com/IsaacdNigerian/status/2118254742245724486
IsaacdNigerian adds a one-line mapping that prevents real damage: an agent loop is a retry loop with a non-deterministic worker, so every tool that writes needs idempotency. A second person in the same thread calls it a retry-with-backoff state machine that happens to call an LLM, which lets old debugging skills apply. Backend engineers already know this failure class. They just need to notice it's the same one.
https://x.com/IsaacdNigerian/status/2118254742245724486
IsaacdNigerian adds a one-line mapping that prevents real damage: an agent loop is a retry loop with a non-deterministic worker, so every tool that writes needs idempotency. A second person in the same thread calls it a retry-with-backoff state machine that happens to call an LLM, which lets old debugging skills apply. Backend engineers already know this failure class. They just need to notice it's the same one.
#26
@lxe
https://x.com/lxe/status/2118760881383489781
lxe asked who is building harnesses for long-horizon auto-research, self-directed optimization and hill-climbing problems, and got 56 likes. A question like that tells you the gap: coding harnesses are everywhere, and harnesses for search problems with a score at the end still aren't. Replies pointed to local tools and a few early projects. The demand is clear, and no default has emerged.
https://x.com/lxe/status/2118760881383489781
lxe asked who is building harnesses for long-horizon auto-research, self-directed optimization and hill-climbing problems, and got 56 likes. A question like that tells you the gap: coding harnesses are everywhere, and harnesses for search problems with a score at the end still aren't. Replies pointed to local tools and a few early projects. The demand is clear, and no default has emerged.
#27
@crazyMLguy
https://x.com/crazyMLguy/status/2118459885205819467
crazyMLguy reports an open source RAG project, PipesHub, ranking first on EnterpriseRAG-Bench with 93.24 overall and 95 percent correctness, ahead of systems from OpenAI, Google, Microsoft and Amazon. The test ran 500 questions over a synthetic corpus of 500,000-plus documents. It combines hybrid search, knowledge graphs and document reading, and an agent loop that searches, inspects results and refines queries. A synthetic corpus deserves some caution. The loop around retrieval, not any single index, is what's doing the work here.
https://x.com/crazyMLguy/status/2118459885205819467
crazyMLguy reports an open source RAG project, PipesHub, ranking first on EnterpriseRAG-Bench with 93.24 overall and 95 percent correctness, ahead of systems from OpenAI, Google, Microsoft and Amazon. The test ran 500 questions over a synthetic corpus of 500,000-plus documents. It combines hybrid search, knowledge graphs and document reading, and an agent loop that searches, inspects results and refines queries. A synthetic corpus deserves some caution. The loop around retrieval, not any single index, is what's doing the work here.
#28
@AnnatarXBT
https://x.com/AnnatarXBT/status/2118236370107871624
AnnatarXBT highlights Harvard's CS2680, Modern AI Systems, which starts at Claude Code, the agent loop, evals and token cost, then goes down through deployment, routing, KV cache, quantization and speculative decoding. All 5 assignments take apart the same agent system: study it, build its harness, then optimize down to inference. Teaching the stack from the agent loop downward says how fast this became core material. The syllabus is worth copying for anyone aiming at agent infrastructure.
https://x.com/AnnatarXBT/status/2118236370107871624
AnnatarXBT highlights Harvard's CS2680, Modern AI Systems, which starts at Claude Code, the agent loop, evals and token cost, then goes down through deployment, routing, KV cache, quantization and speculative decoding. All 5 assignments take apart the same agent system: study it, build its harness, then optimize down to inference. Teaching the stack from the agent loop downward says how fast this became core material. The syllabus is worth copying for anyone aiming at agent infrastructure.
#29
@wallicestene
https://x.com/wallicestene/status/2118769946419249576
wallicestene, who builds scrapers for clients, entered a Meta hackathon with Dapi: say what public data you need, Muse Wonder investigates the site in an agentic loop and gives a strategy, Muse Spark writes and tests the extraction code, and it runs on Apify. Datasets come back as CSV, Excel or JSON with a receipt for every value. Most clients just need the data, the author notes, so the code and investigation were the bottleneck. Per-value receipts are the part that separates this from another scraping wrapper.
https://x.com/wallicestene/status/2118769946419249576
wallicestene, who builds scrapers for clients, entered a Meta hackathon with Dapi: say what public data you need, Muse Wonder investigates the site in an agentic loop and gives a strategy, Muse Spark writes and tests the extraction code, and it runs on Apify. Datasets come back as CSV, Excel or JSON with a receipt for every value. Most clients just need the data, the author notes, so the code and investigation were the bottleneck. Per-value receipts are the part that separates this from another scraping wrapper.
π‘ Eco Products Radar
Eco Products Radar
Autoresearch (Karpathy): the keep-or-discard method everyone cites, plus a wave of one-prompt versions
Claude Code and Opus 5.5: the default runner in most loops today
Inkling 2 Small and Tinker: Thinking Machines' model and fine-tune platform, with autoresearch tests in the launch
Qwen3.8 and Qwen3.5: the local models people tune and run loops on
SpecPilot: agentic tuning of speculative decoding, several summaries today
DeepSeek Harness: replaceable agent loop and a community context visualizer
sfcompute autoresearch: GPU delegation for agent-run training
Grok Bot: named as a loop runner in copied prompts
Autoresearch (Karpathy): the keep-or-discard method everyone cites, plus a wave of one-prompt versions
Claude Code and Opus 5.5: the default runner in most loops today
Inkling 2 Small and Tinker: Thinking Machines' model and fine-tune platform, with autoresearch tests in the launch
Qwen3.8 and Qwen3.5: the local models people tune and run loops on
SpecPilot: agentic tuning of speculative decoding, several summaries today
DeepSeek Harness: replaceable agent loop and a community context visualizer
sfcompute autoresearch: GPU delegation for agent-run training
Grok Bot: named as a loop runner in copied prompts
Comments