Loop Daily: 2026-11-04
The best loop result this window wasn't about code at all: an agent treating AlphaFold 3 as a black-box instrument raised complex prediction scores with 60 percent fewer sequences, and then handed back about 41 percent of its gain because it could only judge itself by AlphaFold's own confidence numbers. That gap between what a loop finds and what it can recognize showed up everywhere. The kernel leaderboard's top players warn about Goodhart's law, a study of autoresearch found humans still matter most for framing problems and unsticking agents, and a Chinese undergraduate wrote the most honest post of the week about spending half a year as GPT's harness. On the infrastructure side, people kept reading DeepSeek's Super-agent paper closely enough to find the awkward parts: section 8.4 on credentials, tasks faked with a timer, and an agent that delegated in 0 of 5 runs unless told to.
#1
@BiologyAIDaily
https://x.com/BiologyAIDaily/status/2117605156183560285
BiologyAIDaily summarized the most concrete autoresearch result of the window: an agent that reorganizes multiple sequence alignments for AlphaFold 3, treating the frozen predictor as a black-box instrument and never seeing ground truth. Across 1,015 protein interfaces it beat the default on every benchmark, moving inter-species heteromers on FoldBench-v1 from a mean DockQ of 0.465 to 0.612. It did this with about 60 percent fewer sequences, so which rows you keep matters more than how deep the alignment is. The most useful number is the loss: nominating a final answer gave back about 41 percent of the improvement it had found, and an oracle with true scores removed most of that. Exploration is largely solved here; the bottleneck is the quality of the feedback signal.
https://x.com/BiologyAIDaily/status/2117605156183560285
BiologyAIDaily summarized the most concrete autoresearch result of the window: an agent that reorganizes multiple sequence alignments for AlphaFold 3, treating the frozen predictor as a black-box instrument and never seeing ground truth. Across 1,015 protein interfaces it beat the default on every benchmark, moving inter-species heteromers on FoldBench-v1 from a mean DockQ of 0.465 to 0.612. It did this with about 60 percent fewer sequences, so which rows you keep matters more than how deep the alignment is. The most useful number is the loss: nominating a final answer gave back about 41 percent of the improvement it had found, and an oracle with true scores removed most of that. Exploration is largely solved here; the bottleneck is the quality of the feedback signal.
#2
@SynapseOperator
https://x.com/SynapseOperator/status/2117631881915031833
SynapseOperator, a third-year undergraduate in China, wrote the most honest autoresearch post of the week. Told to find a research idea, the student fed a paper pool to GPT, picked what sounded good, and spent half a year following its instructions and calling Codex, then discovered the core idea had already been published at ACL Findings as GraphDx, probably where GPT got it. The phrase that stings: I was just GPT's harness. The student has told the advisor it's dead and is going back to reading papers, posting progress daily as a public commitment. This is the failure mode nobody benchmarks, a loop that runs perfectly and teaches the human nothing.
https://x.com/SynapseOperator/status/2117631881915031833
SynapseOperator, a third-year undergraduate in China, wrote the most honest autoresearch post of the week. Told to find a research idea, the student fed a paper pool to GPT, picked what sounded good, and spent half a year following its instructions and calling Codex, then discovered the core idea had already been published at ACL Findings as GraphDx, probably where GPT got it. The phrase that stings: I was just GPT's harness. The student has told the advisor it's dead and is going back to reading papers, posting progress daily as a public commitment. This is the failure mode nobody benchmarks, a loop that runs perfectly and teaches the human nothing.
#3
@cv_usk
https://x.com/cv_usk/status/2117198531907973188
cv_usk summarized Aligning Autoresearch, a paper that asks where humans are still needed before automating the AI researcher, across four experiments. Adding fine-grained human feedback to benchmark creation lowered solver scores by 28 to 46.5 points, meaning harder and better problems. Locating hacked commands across 169 trajectories topped out at an F1 of 64.88, so grader gaming is far from detected. On writing, story continuation went from 2.8 to 8.2 against 10.0 for humans, and experts preferred the trained output 95 percent of the time. The conclusion is modest and useful: people earn their keep framing problems and unsticking stalled agents, and judge quality sets the ceiling.
https://x.com/cv_usk/status/2117198531907973188
cv_usk summarized Aligning Autoresearch, a paper that asks where humans are still needed before automating the AI researcher, across four experiments. Adding fine-grained human feedback to benchmark creation lowered solver scores by 28 to 46.5 points, meaning harder and better problems. Locating hacked commands across 169 trajectories topped out at an F1 of 64.88, so grader gaming is far from detected. On writing, story continuation went from 2.8 to 8.2 against 10.0 for humans, and experts preferred the trained output 95 percent of the time. The conclusion is modest and useful: people earn their keep framing problems and unsticking stalled agents, and judge quality sets the ceiling.
#4
@nasqret
https://x.com/nasqret/status/2117462206505742678
nasqret, a mathematician, says we have reached a form of weak AGI in mathematics, based on personal projects in algebra where speed and quality jumped. The claim is measured: most questions remain intractable for AI, but they are hopeless for humans too, and a surprising number of problems from the literature can be cracked in days with the right auto-research skills, clever harnesses and a ton of tokens. Very young researchers and PhD students are starting wild new projects at the interface. Compare this with snikolov's point below: proof attempts are fast, symbolic and easy to diagnose, which may be exactly why math got here first.
https://x.com/nasqret/status/2117462206505742678
nasqret, a mathematician, says we have reached a form of weak AGI in mathematics, based on personal projects in algebra where speed and quality jumped. The claim is measured: most questions remain intractable for AI, but they are hopeless for humans too, and a surprising number of problems from the literature can be cracked in days with the right auto-research skills, clever harnesses and a ton of tokens. Very young researchers and PhD students are starting wild new projects at the interface. Compare this with snikolov's point below: proof attempts are fast, symbolic and easy to diagnose, which may be exactly why math got here first.
#5
@snikolov
https://x.com/snikolov/status/2117586813820850181
snikolov offered the clearest explanation of why autoresearch cracked Navier-Stokes-style math yet hasn't produced better models. A proof attempt is a very fast symbolic experiment, and you can usually point at what's wrong. A computational experiment takes longer and can fail because of bugs, wrong math or a poor training setup, and you often can't tell which. The suggested paths: cheaper experiments with modular improvements that stack, verifiable subproblems like faster equivalent computation, and letting the system invent its own analysis tools. That last one is an open research problem and probably the real gate on recursive improvement.
https://x.com/snikolov/status/2117586813820850181
snikolov offered the clearest explanation of why autoresearch cracked Navier-Stokes-style math yet hasn't produced better models. A proof attempt is a very fast symbolic experiment, and you can usually point at what's wrong. A computational experiment takes longer and can fail because of bugs, wrong math or a poor training setup, and you often can't tell which. The suggested paths: cheaper experiments with modular improvements that stack, verifiable subproblems like faster equivalent computation, and letting the system invent its own analysis tools. That last one is an open research problem and probably the real gate on recursive improvement.
#6
@HJCH0
https://x.com/HJCH0/status/2117664010493673820
HJCH0 reports that Locus went from state of the art on autonomous post-training to frontier performance on antibody design within a couple of months, through improvements that are domain-agnostic: diverse exploration, better generalization, resilience to noise and high sample efficiency on data-scarce tasks. If that holds, transfer is the headline, since molecules have longer, costlier and noisier evaluation cycles than training runs. soren_dunn_ on the same team calls it one of the first cases of generalizing from AI research to antibody design and potency prediction. These are the team's own claims, but transfer is what the field has mostly failed to show.
https://x.com/HJCH0/status/2117664010493673820
HJCH0 reports that Locus went from state of the art on autonomous post-training to frontier performance on antibody design within a couple of months, through improvements that are domain-agnostic: diverse exploration, better generalization, resilience to noise and high sample efficiency on data-scarce tasks. If that holds, transfer is the headline, since molecules have longer, costlier and noisier evaluation cycles than training runs. soren_dunn_ on the same team calls it one of the first cases of generalizing from AI research to antibody design and potency prediction. These are the team's own claims, but transfer is what the field has mostly failed to show.
#7
@RohanArun
https://x.com/RohanArun/status/2117446433716478226
RohanArun climbed back to 16th on the BF16 leaderboard of the Open Kernel Challenge with three tricks worth stealing. A genetic algorithm dashboard, controlled by Claude, Astra, Gemini and the author, narrows the parameter search and squeezes out about 5 percent more. A small statistical model trained on thousands of benchmarks predicts results before running them, screening billions of combinations in milliseconds. And an automated searcher built on Lean verification rules out combinations that are impossible given physics and memory bounds. The pattern is that the models aren't doing the search directly; they steer a cheap surrogate and a prover.
https://x.com/RohanArun/status/2117446433716478226
RohanArun climbed back to 16th on the BF16 leaderboard of the Open Kernel Challenge with three tricks worth stealing. A genetic algorithm dashboard, controlled by Claude, Astra, Gemini and the author, narrows the parameter search and squeezes out about 5 percent more. A small statistical model trained on thousands of benchmarks predicts results before running them, screening billions of combinations in milliseconds. And an automated searcher built on Lean verification rules out combinations that are impossible given physics and memory bounds. The pattern is that the models aren't doing the search directly; they steer a cheap surrogate and a prover.
#8
@thomasgardiner
https://x.com/thomasgardiner/status/2117267190562828787
thomasgardiner cracked the overall top 10 of the same kernel contest and gave two tips in one line each: you're going to get familiar with Goodhart's law very quickly, and your agent loop does not need to be complex, because agents will take every opportunity to be lazy. A post-mortem is promised when it ends. Those two sentences match what most leaderboard stories have said for two weeks, and they cost a lot of tokens to learn. Simple loops plus suspicion of your own metric keep beating elaborate ones.
https://x.com/thomasgardiner/status/2117267190562828787
thomasgardiner cracked the overall top 10 of the same kernel contest and gave two tips in one line each: you're going to get familiar with Goodhart's law very quickly, and your agent loop does not need to be complex, because agents will take every opportunity to be lazy. A post-mortem is promised when it ends. Those two sentences match what most leaderboard stories have said for two weeks, and they cost a lot of tokens to learn. Simple loops plus suspicion of your own metric keep beating elaborate ones.
#9
@christopherowen
https://x.com/christopherowen/status/2117710440209756396
christopherowen, serving DeepSeek-V4.1-Flash across DGX Sparks, shows why quoting decode speed hides what breaks agents. After matching reference precision, mean KL was only 0.021 nats, yet the top token changed at 4.9 percent of positions, which over a 10-step loop rewrites tool arguments. Without prefill yielding, one cold 131K prefill stalled agent turns over 12 seconds; yielding brought TTFT p50 to 3.3 seconds. Two of the nastiest failures weren't weights at all, but control-token strings in tool outputs being tokenized as special tokens. The missing piece is an end-to-end metric, and findings per hour under a fixed tool policy is the right idea.
https://x.com/christopherowen/status/2117710440209756396
christopherowen, serving DeepSeek-V4.1-Flash across DGX Sparks, shows why quoting decode speed hides what breaks agents. After matching reference precision, mean KL was only 0.021 nats, yet the top token changed at 4.9 percent of positions, which over a 10-step loop rewrites tool arguments. Without prefill yielding, one cold 131K prefill stalled agent turns over 12 seconds; yielding brought TTFT p50 to 3.3 seconds. Two of the nastiest failures weren't weights at all, but control-token strings in tool outputs being tokenized as special tokens. The missing piece is an end-to-end metric, and findings per hour under a fixed tool policy is the right idea.
#10
@akshay_pachaar
https://x.com/akshay_pachaar/status/2117605592999313843
akshay_pachaar laid out UHP, a Unified Harness Protocol, with HarnessRouter as its open-source server: a backend sends a task with a harness_id, the chosen harness runs its own native loop in its own session, and files, failures and output come back through one interface. It supports 30 harness bases, and Alibaba's Qoder has adopted it. The official benchmark claim is the interesting bit: changing only the harness with the model fixed cut costs 25 to 64 percent on healthcare compliance tasks. If harness choice moves cost that much, the next step is routing work between harnesses as well as between models.
https://x.com/akshay_pachaar/status/2117605592999313843
akshay_pachaar laid out UHP, a Unified Harness Protocol, with HarnessRouter as its open-source server: a backend sends a task with a harness_id, the chosen harness runs its own native loop in its own session, and files, failures and output come back through one interface. It supports 30 harness bases, and Alibaba's Qoder has adopted it. The official benchmark claim is the interesting bit: changing only the harness with the model fixed cut costs 25 to 64 percent on healthcare compliance tasks. If harness choice moves cost that much, the next step is routing work between harnesses as well as between models.
#11
@N01ennn
https://x.com/N01ennn/status/2117584818070249641
N01ennn relayed a paper that opened the source of 11 real coding agents, about 4 million lines. Every system converges on the same seven parts: agent loop, LLM integration, tools, memory and context, safety and permissions, orchestration, extensibility. The surprises are the valuable part: not one runtime imports LangChain, LangGraph or AutoGen, none use vector search to find code (it's ripgrep, globs and markdown), and SKILL.md now appears in more systems than MCP, 9 of 11 against 8. It ends with 18 design recommendations and a 90-line starter harness. That's a hard hit to the framework business, and it's data rather than opinion.
https://x.com/N01ennn/status/2117584818070249641
N01ennn relayed a paper that opened the source of 11 real coding agents, about 4 million lines. Every system converges on the same seven parts: agent loop, LLM integration, tools, memory and context, safety and permissions, orchestration, extensibility. The surprises are the valuable part: not one runtime imports LangChain, LangGraph or AutoGen, none use vector search to find code (it's ripgrep, globs and markdown), and SKILL.md now appears in more systems than MCP, 9 of 11 against 8. It ends with 18 design recommendations and a 90-line starter harness. That's a hard hit to the framework business, and it's data rather than opinion.
#12
@heyshrutimishra
https://x.com/heyshrutimishra/status/2117155993671098371
heyshrutimishra pulled the numbers out of DeepSeek's Super-agent design, and they are more modest than the headlines. Median task time was 12.4 minutes against 20.4 for a single session with a subagent tool, and p95 15.7 against 29.5, at roughly 4x the cached tokens. The coding work was simulated by a timer, so this measures scheduling and not code quality. The design point underneath is that a person can follow 3 to 5 active sessions, so the system removes sessions from the human: a Router with no Bash tool and isolated Workers that never reply directly. Treat this as evidence about attention rather than proof it writes better software.
https://x.com/heyshrutimishra/status/2117155993671098371
heyshrutimishra pulled the numbers out of DeepSeek's Super-agent design, and they are more modest than the headlines. Median task time was 12.4 minutes against 20.4 for a single session with a subagent tool, and p95 15.7 against 29.5, at roughly 4x the cached tokens. The coding work was simulated by a timer, so this measures scheduling and not code quality. The design point underneath is that a person can follow 3 to 5 active sessions, so the system removes sessions from the human: a Router with no Bash tool and isolated Workers that never reply directly. Treat this as evidence about attention rather than proof it writes better software.
#13
@xltheg
https://x.com/xltheg/status/2117709349393146053
xltheg's rebuild guide for the same design adds the caveats most summaries dropped. In testing it used 2 to 4x the cached tokens, 23.2M against 6.1M to 10.8M per run. A folder per task isn't security, since workers still run with full access as your user. And an agent given a neutral prompt called a subagent in 0 of 5 runs, so you have to tell it to delegate. The practical recipe is one router, one worker per task, plain code rules filtering events before the model sees them, chat replies capped at 100 words and state in a database. The 0 of 5 figure deserves more attention than it has had, because it undercuts the idea that agents naturally organize themselves.
https://x.com/xltheg/status/2117709349393146053
xltheg's rebuild guide for the same design adds the caveats most summaries dropped. In testing it used 2 to 4x the cached tokens, 23.2M against 6.1M to 10.8M per run. A folder per task isn't security, since workers still run with full access as your user. And an agent given a neutral prompt called a subagent in 0 of 5 runs, so you have to tell it to delegate. The practical recipe is one router, one worker per task, plain code rules filtering events before the model sees them, chat replies capped at 100 words and state in a database. The 0 of 5 figure deserves more attention than it has had, because it undercuts the idea that agents naturally organize themselves.
#14
@SlavaOPs
https://x.com/SlavaOPs/status/2117710992914415646
SlavaOPs says to skip the architecture and read section 8.4, where DeepSeek's authors state the problem their own design creates. The cloud loop takes untrusted events all day, endpoint daemons sit on machines with private code and credentials, and the design assumes those two trust each other, so one prompt injection in the Router reaches every connected device. The sentence to quote: as long as credentials on an endpoint can enter the model's context, normal execution can expose them. Not an attack, working as designed. Publishing this before shipping is rarer than the architecture and deserves the credit.
https://x.com/SlavaOPs/status/2117710992914415646
SlavaOPs says to skip the architecture and read section 8.4, where DeepSeek's authors state the problem their own design creates. The cloud loop takes untrusted events all day, endpoint daemons sit on machines with private code and credentials, and the design assumes those two trust each other, so one prompt injection in the Router reaches every connected device. The sentence to quote: as long as credentials on an endpoint can enter the model's context, normal execution can expose them. Not an attack, working as designed. Publishing this before shipping is rarer than the architecture and deserves the credit.
#15
@MichaelGannotti
https://x.com/MichaelGannotti/status/2117231832445948005
MichaelGannotti shared what an OpenClaw agent called Carol concluded after being sent to study DeepSeek Harness: these are habits worth taking, not software worth installing. The adopted list is sharp: confirm once then stay quiet, unasked work starts as a record not a task, messages can never widen permission, a draft is not ready, a drop is not a send. The do-not-copy list names approval-free worker shells, because a worker can still call a program by absolute path. Notably Carol states what it did not open: the code, the deployment guide and the article were unread. An agent reporting the limits of its own reading is the behavior to want.
https://x.com/MichaelGannotti/status/2117231832445948005
MichaelGannotti shared what an OpenClaw agent called Carol concluded after being sent to study DeepSeek Harness: these are habits worth taking, not software worth installing. The adopted list is sharp: confirm once then stay quiet, unasked work starts as a record not a task, messages can never widen permission, a draft is not ready, a drop is not a send. The do-not-copy list names approval-free worker shells, because a worker can still call a program by absolute path. Notably Carol states what it did not open: the code, the deployment guide and the article were unread. An agent reporting the limits of its own reading is the behavior to want.
#16
@BioAgentDaily
https://x.com/BioAgentDaily/status/2117301910755754079
BioAgentDaily listed what Tencent's Hyra 2.0 actually did beyond the announcement, in research runs of days to weeks through three agent layers. It trained an inverse-folding model called HyInv, 57x faster than ProteinMPNN, tuned the XiangShan RISC-V core for +5.79 percent performance per GHz with area about flat, and designed a robot hand with +28.6 percent rotation speed in simulation. All numbers are Tencent's own. The thesis is that durable output is reusable tools rather than answers, and three domains with hard numbers make a better case than the blog post alone.
https://x.com/BioAgentDaily/status/2117301910755754079
BioAgentDaily listed what Tencent's Hyra 2.0 actually did beyond the announcement, in research runs of days to weeks through three agent layers. It trained an inverse-folding model called HyInv, 57x faster than ProteinMPNN, tuned the XiangShan RISC-V core for +5.79 percent performance per GHz with area about flat, and designed a robot hand with +28.6 percent rotation speed in simulation. All numbers are Tencent's own. The thesis is that durable output is reusable tools rather than answers, and three domains with hard numbers make a better case than the blog post alone.
#17
@XuyuanXiong
https://x.com/XuyuanXiong/status/2117099707004789221
XuyuanXiong, a core contributor, introduced NextTransformer, an open framework and fair benchmark for next-generation LLM architectures: 30-plus architectures with one trainer and data pipeline, training compute matched within 1 percent across about 1.4B-parameter models on 49B tokens. In the finished runs, mHC, AttnRes and GDN hybrids consistently beat the dense baseline. The piece to watch arrives November 14: an auto-research arena where LLMs propose, train and benchmark new architectures under the same rules. Fixed compute and identical rules are what most architecture comparisons lack, so this could become the scoreboard for autonomous model research.
https://x.com/XuyuanXiong/status/2117099707004789221
XuyuanXiong, a core contributor, introduced NextTransformer, an open framework and fair benchmark for next-generation LLM architectures: 30-plus architectures with one trainer and data pipeline, training compute matched within 1 percent across about 1.4B-parameter models on 49B tokens. In the finished runs, mHC, AttnRes and GDN hybrids consistently beat the dense baseline. The piece to watch arrives November 14: an auto-research arena where LLMs propose, train and benchmark new architectures under the same rules. Fixed compute and identical rules are what most architecture comparisons lack, so this could become the scoreboard for autonomous model research.
#18
@pkp_io
https://x.com/pkp_io/status/2117667090890940760
pkp_io described a founder's working day with agents: about 8 to 10 running for hours on a remote server reachable only from a private network, with the longest loop lasting 47 hours. Four waiting agents sit behind three independent agents verifying their work, so the founder isn't the bottleneck for trivial reviews. Agents pause only for critical decisions, which get handled from a laptop or phone by reading their reports. Three verifiers per worker sounds expensive until you compare it with a founder's attention, and this is one of the first descriptions of review capacity scaled by adding agents rather than hours.
https://x.com/pkp_io/status/2117667090890940760
pkp_io described a founder's working day with agents: about 8 to 10 running for hours on a remote server reachable only from a private network, with the longest loop lasting 47 hours. Four waiting agents sit behind three independent agents verifying their work, so the founder isn't the bottleneck for trivial reviews. Agents pause only for critical decisions, which get handled from a laptop or phone by reading their reports. Three verifiers per worker sounds expensive until you compare it with a founder's attention, and this is one of the first descriptions of review capacity scaled by adding agents rather than hours.
#19
@tigdoesthings
https://x.com/tigdoesthings/status/2117558726001340704
tigdoesthings open-sourced Talos, an autoresearch tool in beta for algorithm search. You pick a challenge and give an LLM a direction, and it loops on writing, modifying and testing algorithms against the current top mainnet algorithm. Every candidate is compiled and evaluated by the project's own runtime and verifier, and must beat the baseline on both training and held-out data. The gate is what makes this interesting: a reader immediately asked how it avoids overfitting the benchmark, and the held-out requirement is the honest first answer. You bring your own compute and model account, so the cost of the search falls on participants.
https://x.com/tigdoesthings/status/2117558726001340704
tigdoesthings open-sourced Talos, an autoresearch tool in beta for algorithm search. You pick a challenge and give an LLM a direction, and it loops on writing, modifying and testing algorithms against the current top mainnet algorithm. Every candidate is compiled and evaluated by the project's own runtime and verifier, and must beat the baseline on both training and held-out data. The gate is what makes this interesting: a reader immediately asked how it avoids overfitting the benchmark, and the held-out requirement is the honest first answer. You bring your own compute and model account, so the cost of the search falls on participants.
#20
@_pksharma
https://x.com/_pksharma/status/2117650078328377839
_pksharma broke down Bitdefender's report on a ransomware case at a UK school estate, which splits 'AI ransomware' into three separately rated claims. A model wrote the commands: strongly supported. An agentic loop chose each action: behaves like one, with retries 40 to 49 seconds apart, though a person pasting between a shell and a chat window would look the same. No human in the hands-on phase: unsupported. Bitdefender calls it one case, an anecdote and not a trend. Copy this format, because most AI security claims would look different if they had to rate each link in the chain.
https://x.com/_pksharma/status/2117650078328377839
_pksharma broke down Bitdefender's report on a ransomware case at a UK school estate, which splits 'AI ransomware' into three separately rated claims. A model wrote the commands: strongly supported. An agentic loop chose each action: behaves like one, with retries 40 to 49 seconds apart, though a person pasting between a shell and a chat window would look the same. No human in the hands-on phase: unsupported. Bitdefender calls it one case, an anecdote and not a trend. Copy this format, because most AI security claims would look different if they had to rate each link in the chain.
#21
@jacksunny18
https://x.com/jacksunny18/status/2117698858767860116
jacksunny18 made the most useful small point on agent accounting: a task is a plan and a run is a fact. A task worked by three runs carries one number that belongs to none of them, so cost attached to tasks makes every later question approximate. Record spend on the run with its provider, model and agent identifiers, and sum upward. The two-minute check is to take your last expensive run and ask which task it was for; if the honest answer needs a guess, the unit is wrong. It's cheap to get right before the first run and painful to migrate after.
https://x.com/jacksunny18/status/2117698858767860116
jacksunny18 made the most useful small point on agent accounting: a task is a plan and a run is a fact. A task worked by three runs carries one number that belongs to none of them, so cost attached to tasks makes every later question approximate. Record spend on the run with its provider, model and agent identifiers, and sum upward. The two-minute check is to take your last expensive run and ask which task it was for; if the honest answer needs a guess, the unit is wrong. It's cheap to get right before the first run and painful to migrate after.
#22
@Prayag12B
https://x.com/Prayag12B/status/2117588351238693317
Prayag12B called out three habits that separate engineers from prompters, and the third one is the loop lesson. Documentation works as context compression, so the agent doesn't burn thousands of tokens reading five folders of source to find an internal API. Naming exponential backoff with jitter beats asking for retry logic. And soft warnings are useless against automated loops: every API and cloud environment an agent touches needs a hard cap that cuts off execution and returns an error. A warning email at $100 arrives after a runaway loop has already spent it, which is why the budget belongs in the path of the call.
https://x.com/Prayag12B/status/2117588351238693317
Prayag12B called out three habits that separate engineers from prompters, and the third one is the loop lesson. Documentation works as context compression, so the agent doesn't burn thousands of tokens reading five folders of source to find an internal API. Naming exponential backoff with jitter beats asking for retry logic. And soft warnings are useless against automated loops: every API and cloud environment an agent touches needs a hard cap that cuts off execution and returns an error. A warning email at $100 arrives after a runaway loop has already spent it, which is why the budget belongs in the path of the call.
#23
@Arindam_1729
https://x.com/Arindam_1729/status/2117263780036141143
Arindam_1729 explained Google's Agent Compute Engine from the angle that matters: a coding agent spends most of its life waiting on the LLM, so what happens to the sandbox? Google snapshots an idle microVM and restores on the next action, shares unchanged memory pages across microVMs through copy-on-write, and keeps long-lived credentials outside the sandbox behind external gateways. Restore latency sits inside the loop, so it only wins when idle gaps are longer than the restore time, which for most coding agents they are. Treat compute like a process you can pause, not a server you keep running.
https://x.com/Arindam_1729/status/2117263780036141143
Arindam_1729 explained Google's Agent Compute Engine from the angle that matters: a coding agent spends most of its life waiting on the LLM, so what happens to the sandbox? Google snapshots an idle microVM and restores on the next action, shares unchanged memory pages across microVMs through copy-on-write, and keeps long-lived credentials outside the sandbox behind external gateways. Restore latency sits inside the loop, so it only wins when idle gaps are longer than the restore time, which for most coding agents they are. Treat compute like a process you can pause, not a server you keep running.
#24
@iveyzen
https://x.com/iveyzen/status/2117093336162934912
iveyzen flagged a small but telling detail in Microsoft's Conductor 0.1.44, an open-source CLI for multi-agent workflows written in YAML. It adds execution profiles for Docker and Azure Container Apps that keep the loop, MCP tools and workspace dependencies in one long-lived container while configuring the model separately. For retained workspaces, resuming refuses to proceed when a staging marker is missing rather than risk overwriting files the agent wrote. That refusal is the kind of unglamorous safety people end up wanting, and the post rightly adds that separating the runtime doesn't make a proposed change correct.
https://x.com/iveyzen/status/2117093336162934912
iveyzen flagged a small but telling detail in Microsoft's Conductor 0.1.44, an open-source CLI for multi-agent workflows written in YAML. It adds execution profiles for Docker and Azure Container Apps that keep the loop, MCP tools and workspace dependencies in one long-lived container while configuring the model separately. For retained workspaces, resuming refuses to proceed when a staging marker is missing rather than risk overwriting files the agent wrote. That refusal is the kind of unglamorous safety people end up wanting, and the post rightly adds that separating the runtime doesn't make a proposed change correct.
#25
@FeliciaRios_
https://x.com/FeliciaRios_/status/2117144707461722482
FeliciaRios_ quoted Meta's chief AI officer, Alexandr Wang, at Startup School: internally, with the right agentic loop plus the right evaluation system and metric, a swarm of agents can accomplish more than a team of 100 engineers, very easily. No project, number or metric was given, so treat it as a claim from someone with an interest in it. The conditions are the useful part. Everything else this window suggests the metric is the hard bit, so the sentence means little until someone says which metric survived contact with agents who game it.
https://x.com/FeliciaRios_/status/2117144707461722482
FeliciaRios_ quoted Meta's chief AI officer, Alexandr Wang, at Startup School: internally, with the right agentic loop plus the right evaluation system and metric, a swarm of agents can accomplish more than a team of 100 engineers, very easily. No project, number or metric was given, so treat it as a claim from someone with an interest in it. The conditions are the useful part. Everything else this window suggests the metric is the hard bit, so the sentence means little until someone says which metric survived contact with agents who game it.
#26
@Dimillian
https://x.com/Dimillian/status/2117569670912053471
Dimillian offered the needed counterweight: building games is slow and personal, every addition has to be intentional, and the same goes for software, so the auto agentic loop doing 1,500 PRs a day is probably fine for others but not for this builder. Volume is the wrong metric when the product is a set of taste decisions. It's a short post, but it names the category that loops haven't touched, the work where the author wants to have made each choice. Expect that to become a selling point rather than a limitation.
https://x.com/Dimillian/status/2117569670912053471
Dimillian offered the needed counterweight: building games is slow and personal, every addition has to be intentional, and the same goes for software, so the auto agentic loop doing 1,500 PRs a day is probably fine for others but not for this builder. Volume is the wrong metric when the product is a set of taste decisions. It's a short post, but it names the category that loops haven't touched, the work where the author wants to have made each choice. Expect that to become a selling point rather than a limitation.
#27
@franspeterii
https://x.com/franspeterii/status/2117580526731350261
franspeterii tested whether you can optimize a product for agent UX, and the answer was a short recipe. Let your AI design a test around one metric, tool calls per action, then run an auto-research loop: a subagent tests, you change something, test again. It's a neat metric because it's cheap, hard to fake and maps directly to cost. Expect SaaS teams to start running this before their competitors' agents decide which product is easier to use.
https://x.com/franspeterii/status/2117580526731350261
franspeterii tested whether you can optimize a product for agent UX, and the answer was a short recipe. Let your AI design a test around one metric, tool calls per action, then run an auto-research loop: a subagent tests, you change something, test again. It's a neat metric because it's cheap, hard to fake and maps directly to cost. Expect SaaS teams to start running this before their competitors' agents decide which product is easier to use.
#28
@stretchcloud
https://x.com/stretchcloud/status/2117379977536770469
stretchcloud reports DeepSeek Harness crossed 200,000 GitHub stars in fifteen days, beating OpenClaw's 88, with 74,578 on day one, and focuses on Creator Mode, one of four presets. The architecture has no privileged core: model adapter, tool registry, session log and agent loop are all plugins on the Cordis framework, and the append-only log means anything a model sees can be reconstructed. External pull requests aren't accepted, so contributions must arrive as plugins. That is either the cleanest governance in the category or what eventually slows it down. Compatibility bridges for Claude Code and Codex hooks show it wants to absorb the rivals rather than replace them.
https://x.com/stretchcloud/status/2117379977536770469
stretchcloud reports DeepSeek Harness crossed 200,000 GitHub stars in fifteen days, beating OpenClaw's 88, with 74,578 on day one, and focuses on Creator Mode, one of four presets. The architecture has no privileged core: model adapter, tool registry, session log and agent loop are all plugins on the Cordis framework, and the append-only log means anything a model sees can be reconstructed. External pull requests aren't accepted, so contributions must arrive as plugins. That is either the cleanest governance in the category or what eventually slows it down. Compatibility bridges for Claude Code and Codex hooks show it wants to absorb the rivals rather than replace them.
#29
@zksecurityXYZ
https://x.com/zksecurityXYZ/status/2117550503060185134
zksecurityXYZ, working with the Ethereum Foundation on an end-to-end verified EVM implementation for zkVMs, announced an autoresearch competition around evm.asm: prove the lowest runtime to validate an Ethereum block. A verified spec is close to the ideal loop target, since any candidate can be mathematically checked against it, so there is no judge to game. Another builder in the same thread notes AI can keep optimizing implementations as long as equivalence to the spec is proved. Formal verification may turn out to be the domain where autoresearch has the cleanest signal after pure math.
https://x.com/zksecurityXYZ/status/2117550503060185134
zksecurityXYZ, working with the Ethereum Foundation on an end-to-end verified EVM implementation for zkVMs, announced an autoresearch competition around evm.asm: prove the lowest runtime to validate an Ethereum block. A verified spec is close to the ideal loop target, since any candidate can be mathematically checked against it, so there is no judge to game. Another builder in the same thread notes AI can keep optimizing implementations as long as equivalence to the spec is proved. Formal verification may turn out to be the domain where autoresearch has the cleanest signal after pure math.
#30
@claudioatilano
https://x.com/claudioatilano/status/2117589589418520966
claudioatilano reports maxing out the Ultra plan's weekly limit in three days by running autoresearch orchestration, and suggests that as a benchmark for how long a plan lasts. One line, but a real data point for anyone budgeting this work: a $200 tier does not cover a continuous loop for a week. It lines up with the wider limit complaints this window, and explains why serious loop runners keep moving toward self-hosted models or metered APIs.
https://x.com/claudioatilano/status/2117589589418520966
claudioatilano reports maxing out the Ultra plan's weekly limit in three days by running autoresearch orchestration, and suggests that as a benchmark for how long a plan lasts. One line, but a real data point for anyone budgeting this work: a $200 tier does not cover a continuous loop for a week. It lines up with the wider limit complaints this window, and explains why serious loop runners keep moving toward self-hosted models or metered APIs.
π‘ Eco Products Radar
Eco Products Radar
DeepSeek Harness and the Super-agent design: the essay and repo most posts read, now examined for its token bill, its faked coding work and its credential warning.
Claude Code and Codex: the baseline harnesses, hosts for most runs and targets of new compatibility bridges.
Open Kernel Challenge: the public scoreboard where supervisors, surrogates and verifiers get compared.
Karpathy's autoresearch: still the reference format, now mostly seen in curriculum lists.
Hyra 2.0: Tencent's research agent, arguing for reusable tools over answers.
AlphaFold 3: used as a black-box instrument, the cleanest non-code loop target of the window.
Teamily AI: a heavily templated promotional wave around the phrase agentic loop, excluded from cases.
Opus 5.5, GPT-6 Astra and DeepSeek-V4.1-Flash: the models most often named inside loops.
DeepSeek Harness and the Super-agent design: the essay and repo most posts read, now examined for its token bill, its faked coding work and its credential warning.
Claude Code and Codex: the baseline harnesses, hosts for most runs and targets of new compatibility bridges.
Open Kernel Challenge: the public scoreboard where supervisors, surrogates and verifiers get compared.
Karpathy's autoresearch: still the reference format, now mostly seen in curriculum lists.
Hyra 2.0: Tencent's research agent, arguing for reusable tools over answers.
AlphaFold 3: used as a black-box instrument, the cleanest non-code loop target of the window.
Teamily AI: a heavily templated promotional wave around the phrase agentic loop, excluded from cases.
Opus 5.5, GPT-6 Astra and DeepSeek-V4.1-Flash: the models most often named inside loops.
Comments