Loop Daily: 2026-09-03
The autoresearch scene had its most structural week yet: a dedicated benchmark for research loops launched and immediately showed stark gaps between models that coding benchmarks no longer reveal, EvoMap open-sourced a workflow whose whole point is that the model that wrote an experiment never grades it, and Fable 5.1 arrived claiming state of the art on autonomous science. The vibe shift is visible in what people argue about: not whether loops work, but who audits them.
#1
@_potatodonkey_
https://x.com/_potatodonkey_/status/2095178982442602545
Autoresearch Bench launched: a benchmark for coding agents autonomously tackling research problems in closed-loop experimental cycles. The pitch lands because ordinary coding benchmarks are saturating while the spread between models inside research loops is still stark. Early results reported around the launch have Opus 5 best overall and Grok 4.6 second and most token-efficient, with tasks being open-sourced. Someone finally built the scoreboard for the thing everyone spent the summer doing informally.
https://x.com/_potatodonkey_/status/2095178982442602545
Autoresearch Bench launched: a benchmark for coding agents autonomously tackling research problems in closed-loop experimental cycles. The pitch lands because ordinary coding benchmarks are saturating while the spread between models inside research loops is still stark. Early results reported around the launch have Opus 5 best overall and Grok 4.6 second and most token-efficient, with tasks being open-sourced. Someone finally built the scoreboard for the thing everyone spent the summer doing informally.
#2
@N01ennn
https://x.com/N01ennn/status/2094848612433875061
The best writeup of EvoMap's newly open-sourced AutoResearch goes straight at the failure nobody says out loud: code running does not mean the research is complete. The workflow separates the seat that produces from the seat that judges: three models invent ideas independently and review each other, a cheap pilot runs first, a failed pilot stays on disk as a real result, the model that wrote the experiment never grades it, and a blind reviewer challenges the claim before the engine, not the agent, closes the project. His framing is exactly right: a single model grading its own work is the same distribution asked twice, and this is the first agent stack that copies peer review's org chart instead of its vibe.
https://x.com/N01ennn/status/2094848612433875061
The best writeup of EvoMap's newly open-sourced AutoResearch goes straight at the failure nobody says out loud: code running does not mean the research is complete. The workflow separates the seat that produces from the seat that judges: three models invent ideas independently and review each other, a cheap pilot runs first, a failed pilot stays on disk as a real result, the model that wrote the experiment never grades it, and a blind reviewer challenges the claim before the engine, not the agent, closes the project. His framing is exactly right: a single model grading its own work is the same distribution asked twice, and this is the first agent stack that copies peer review's org chart instead of its vibe.
#3
@denixbt
https://x.com/denixbt/status/2095140684059598954
The skeptical read of the same AutoResearch paper, and it is the checklist to keep: how many times did a human step in, how many repeats and how much does the number wobble, who ran the audit, and what was the funnel. The honest numbers in the paper: one 8-GPU server for a week produced roughly 2,584 ideas, 355 survived review, 22 became experiments, 14 held up. One idea in two hundred. His verdict of promising, not proven is fair, and the four questions apply to every autonomous-discovery press release from now on.
https://x.com/denixbt/status/2095140684059598954
The skeptical read of the same AutoResearch paper, and it is the checklist to keep: how many times did a human step in, how many repeats and how much does the number wobble, who ran the audit, and what was the funnel. The honest numbers in the paper: one 8-GPU server for a week produced roughly 2,584 ideas, 355 survived review, 22 became experiments, 14 held up. One idea in two hundred. His verdict of promising, not proven is fair, and the four questions apply to every autonomous-discovery press release from now on.
#4
@j_foerst
https://x.com/j_foerst/status/2094738793781710951
Jakob Foerster's group tackled the next bottleneck: autoresearch is great when experiments are fast and cheap, but what happens when a run costs weeks and millions? Human scientists get around this with small-scale proofs of concept and scaling curves, and this work is a step toward giving agents the same judgment: choosing what experiment to run given the cost and accuracy trade-off curve. This is the research direction that decides whether loops stay a hobby for 5-minute training runs or graduate to real science budgets.
https://x.com/j_foerst/status/2094738793781710951
Jakob Foerster's group tackled the next bottleneck: autoresearch is great when experiments are fast and cheap, but what happens when a run costs weeks and millions? Human scientists get around this with small-scale proofs of concept and scaling curves, and this work is a step toward giving agents the same judgment: choosing what experiment to run given the cost and accuracy trade-off curve. This is the research direction that decides whether loops stay a hobby for 5-minute training runs or graduate to real science budgets.
#5
@askalphaxiv
https://x.com/askalphaxiv/status/2094854592995725324
Fable 5.1 is now state of the art for autoresearch: 52.6 percent on Terminal-Bench-Science against 24.7 for Fable 5 and 22.4 for GPT 5.6 Sol, more than doubling its predecessor. alphaXiv wired it into OpenResearch the same day. Model releases are now marketed on autonomous-research capability first; that itself is the news.
https://x.com/askalphaxiv/status/2094854592995725324
Fable 5.1 is now state of the art for autoresearch: 52.6 percent on Terminal-Bench-Science against 24.7 for Fable 5 and 22.4 for GPT 5.6 Sol, more than doubling its predecessor. alphaXiv wired it into OpenResearch the same day. Model releases are now marketed on autonomous-research capability first; that itself is the news.
#6
@anabology
https://x.com/anabology/status/2095151075544236464
Autoresearch left the GPU cluster: an overnight Fable 5.1 optimization run on MRI physics, tuning the mapping from bitmap to time-averaged water magnetization. Medical imaging parameter search as an unattended loop is precisely the kind of domain transfer the method promised, and it is happening in individual labs, not just at model companies.
https://x.com/anabology/status/2095151075544236464
Autoresearch left the GPU cluster: an overnight Fable 5.1 optimization run on MRI physics, tuning the mapping from bitmap to time-averaged water magnetization. Medical imaging parameter search as an unattended loop is precisely the kind of domain transfer the method promised, and it is happening in individual labs, not just at model companies.
#7
@PMocz
https://x.com/PMocz/status/2095122198243537377
A physics demo where the agent discovers the governing physics of a black-box simulator on its own, from a researcher who works on scientific computing. His extrapolation is the interesting part: imagine a properly scaled research institute of agents with layers of physics verifiers and native orchestration. The verifier layer keeps showing up as the missing piece in every serious proposal.
https://x.com/PMocz/status/2095122198243537377
A physics demo where the agent discovers the governing physics of a black-box simulator on its own, from a researcher who works on scientific computing. His extrapolation is the interesting part: imagine a properly scaled research institute of agents with layers of physics verifiers and native orchestration. The verifier layer keeps showing up as the missing piece in every serious proposal.
#8
@theotherpomp
https://x.com/theotherpomp/status/2095259048337625539
A working example of loop-on-loop composition: using Karpathy's autoresearch harness with Fable driving, to optimize GLM 5.3 Flash tensor-parallel inference, starting from a community recipe. Cheap open models tuned by expensive frontier models inside an automated loop is a real cost structure now, not a thought experiment.
https://x.com/theotherpomp/status/2095259048337625539
A working example of loop-on-loop composition: using Karpathy's autoresearch harness with Fable driving, to optimize GLM 5.3 Flash tensor-parallel inference, starting from a community recipe. Cheap open models tuned by expensive frontier models inside an automated loop is a real cost structure now, not a thought experiment.
#9
@0xNeoArch
https://x.com/0xNeoArch/status/2095138820983308414
The honest failure report: he had Fable 5.1 run autoresearch for over eight hours building a custom quantized-inference recipe and his GPUs still will not break 200 watts. Eight hours of autonomous experimentation producing a whole recipe that does not fix the bottleneck is what most loops actually look like, and posting it as a question rather than a victory lap is more useful than half the success stories.
https://x.com/0xNeoArch/status/2095138820983308414
The honest failure report: he had Fable 5.1 run autoresearch for over eight hours building a custom quantized-inference recipe and his GPUs still will not break 200 watts. Eight hours of autonomous experimentation producing a whole recipe that does not fix the bottleneck is what most loops actually look like, and posting it as a question rather than a victory lap is more useful than half the success stories.
#10
@AutoTrustAI
https://x.com/AutoTrustAI/status/2094739630314889235
AutoTrust's ScienceGuru harness running their Guru Turbo model took first place on Autoresearch@Home with a val_bpb of 0.889522, and the interesting claim is upstream: the same hypothesis-code-experiment-audit loop supervised the model's own post-training, which they say compressed an 800B model's post-training from quarters to weeks. Self-referential, unverified, and exactly the kind of claim Autoresearch@Home exists to make public and reproducible.
https://x.com/AutoTrustAI/status/2094739630314889235
AutoTrust's ScienceGuru harness running their Guru Turbo model took first place on Autoresearch@Home with a val_bpb of 0.889522, and the interesting claim is upstream: the same hypothesis-code-experiment-audit loop supervised the model's own post-training, which they say compressed an 800B model's post-training from quarters to weeks. Self-referential, unverified, and exactly the kind of claim Autoresearch@Home exists to make public and reproducible.
#11
@stretchcloud
https://x.com/stretchcloud/status/2094297380409811113
OpenResearch CLI runs the entire autoresearch loop locally: Claude Code, Codex and OpenCode working in parallel isolated worktrees on literature review, experiment analysis and synthesis, with data never leaving the machine. He built Campfire as the general coordination layer on top: majority-rules permission voting, session replay, agent races across worktrees, shared semantic memory. The gap he names is real: most people running multiple agent CLIs are still stitching outputs together by hand.
https://x.com/stretchcloud/status/2094297380409811113
OpenResearch CLI runs the entire autoresearch loop locally: Claude Code, Codex and OpenCode working in parallel isolated worktrees on literature review, experiment analysis and synthesis, with data never leaving the machine. He built Campfire as the general coordination layer on top: majority-rules permission voting, session replay, agent races across worktrees, shared semantic memory. The gap he names is real: most people running multiple agent CLIs are still stitching outputs together by hand.
#12
@imfabiokeller
https://x.com/imfabiokeller/status/2095251245212516398
A practitioner postmortem worth more than most papers: his app's agent loop checked every acceptance criterion and passed all 4,600 tests, and the feature did not work at all. Root causes: parallel subagents each got a reviewer, but reviewers deferred judgment because other agents' parts were missing, so now he only parallelizes work that follows a single user path; and UI needed its own iteration skill where the orchestrator actually drives a browser and compares against mockups pixel by pixel. Green tests are not evidence; this is the small-scale version of the same lesson EvoMap institutionalized.
https://x.com/imfabiokeller/status/2095251245212516398
A practitioner postmortem worth more than most papers: his app's agent loop checked every acceptance criterion and passed all 4,600 tests, and the feature did not work at all. Root causes: parallel subagents each got a reviewer, but reviewers deferred judgment because other agents' parts were missing, so now he only parallelizes work that follows a single user path; and UI needed its own iteration skill where the orchestrator actually drives a browser and compares against mockups pixel by pixel. Green tests are not evidence; this is the small-scale version of the same lesson EvoMap institutionalized.
#13
@IsaacWalde20269
https://x.com/IsaacWalde20269/status/2094943922535403925
A one-man marketing shop turned a daily ads report into a proper agentic loop: ad, email, organic and landing-page behavior data lands in a private GitHub repo, chained skills analyze across fragments with one goal, raise ROAS to a 60 target, the agent proposes adjustments and then drives the ads manager itself, and a closing /ads-loop verifies the changes were applied correctly. Fetch, decide, act, verify, all outside a dashboard he never opens. Small business operations is quietly becoming the largest autoresearch domain.
https://x.com/IsaacWalde20269/status/2094943922535403925
A one-man marketing shop turned a daily ads report into a proper agentic loop: ad, email, organic and landing-page behavior data lands in a private GitHub repo, chained skills analyze across fragments with one goal, raise ROAS to a 60 target, the agent proposes adjustments and then drives the ads manager itself, and a closing /ads-loop verifies the changes were applied correctly. Fetch, decide, act, verify, all outside a dashboard he never opens. Small business operations is quietly becoming the largest autoresearch domain.
#14
@serge_ai_lab
https://x.com/serge_ai_lab/status/2095183469882290336
The economics insight of the week from the DeepSWE board: Gemini 3.8 Flash hit 74 percent by taking 166 steps for 2.36 dollars, while Fable 5 scored 70 at 21.63 dollars in 88 steps. Cheap tokens mean the model can afford to be wrong, inspect the error, and retry 166 times. Cheap tokens plus deep agentic loops beating expensive monolithic models is a strategy statement, and it explains why every lab is suddenly competing on cost per loop iteration rather than single-shot IQ.
https://x.com/serge_ai_lab/status/2095183469882290336
The economics insight of the week from the DeepSWE board: Gemini 3.8 Flash hit 74 percent by taking 166 steps for 2.36 dollars, while Fable 5 scored 70 at 21.63 dollars in 88 steps. Cheap tokens mean the model can afford to be wrong, inspect the error, and retry 166 times. Cheap tokens plus deep agentic loops beating expensive monolithic models is a strategy statement, and it explains why every lab is suddenly competing on cost per loop iteration rather than single-shot IQ.
#15
@cursor_ai
https://x.com/cursor_ai/status/2095257412781396114
Cursor shipped self-hosted machines for cloud agents: the agent loop stays in Cursor while tool execution, code, build outputs and secrets stay on infrastructure you manage, including autoscaling pools, with E2B, Modal, Cloudflare and others as sandbox partners. The architecture split everyone converged on this week, loop in the vendor's cloud, hands inside your network, is now a first-class product, and it is what lets regulated teams run long loops at all.
https://x.com/cursor_ai/status/2095257412781396114
Cursor shipped self-hosted machines for cloud agents: the agent loop stays in Cursor while tool execution, code, build outputs and secrets stay on infrastructure you manage, including autoscaling pools, with E2B, Modal, Cloudflare and others as sandbox partners. The architecture split everyone converged on this week, loop in the vendor's cloud, hands inside your network, is now a first-class product, and it is what lets regulated teams run long loops at all.
#16
@vijpatel7
https://x.com/vijpatel7/status/2095188597427122511
A small but telling build: a QA agent whose only job is to adversarially test his voice agents, produce transcripts and recordings, and feed improvements back into the voice agent from the results. Agents whose entire purpose is to attack other agents and close the improvement loop are becoming a standard team seat.
https://x.com/vijpatel7/status/2095188597427122511
A small but telling build: a QA agent whose only job is to adversarially test his voice agents, produce transcripts and recordings, and feed improvements back into the voice agent from the results. Agents whose entire purpose is to attack other agents and close the improvement loop are becoming a standard team seat.
π‘ Eco Products Radar
Eco Products Radar
autoresearch (Karpathy) - the program.md-driven training-optimization harness, now the generic verb for the whole practice
AutoResearch (EvoMap) - newly open-sourced research workflow with producer-judge separation and blind review
Autoresearch Bench - the new benchmark for closed-loop research agents; Opus 5 leads, Grok 4.6 most efficient
Autoresearch@Home - ensue_ai's public leaderboard for reproducible self-improvement runs
OpenResearch - alphaXiv's hosted research-loop product, first to ship Fable 5.1
Fable 5.1 - Anthropic's new model, doubling Terminal-Bench-Science and instantly the default autoresearch engine
Gemini 3.8 Flash - Google's cheap-token deep-loop play, self-correcting agentic demos and top DeepSWE score
Cursor Cloud Agents - self-hosted machines split the loop from the execution boundary
Claude Code / Codex / OpenCode - the trio composed into local research loops by OpenResearch CLI and Campfire
autoresearch (Karpathy) - the program.md-driven training-optimization harness, now the generic verb for the whole practice
AutoResearch (EvoMap) - newly open-sourced research workflow with producer-judge separation and blind review
Autoresearch Bench - the new benchmark for closed-loop research agents; Opus 5 leads, Grok 4.6 most efficient
Autoresearch@Home - ensue_ai's public leaderboard for reproducible self-improvement runs
OpenResearch - alphaXiv's hosted research-loop product, first to ship Fable 5.1
Fable 5.1 - Anthropic's new model, doubling Terminal-Bench-Science and instantly the default autoresearch engine
Gemini 3.8 Flash - Google's cheap-token deep-loop play, self-correcting agentic demos and top DeepSWE score
Cursor Cloud Agents - self-hosted machines split the loop from the execution boundary
Claude Code / Codex / OpenCode - the trio composed into local research loops by OpenResearch CLI and Campfire
Comments