Loop Daily: August 6, 2026
The loop conversation split into two clear camps today. On one side, working loops shipping real artifacts: a GPS-free drone navigation model invented from scratch in 81 iterations, four upstream performance PRs to Zod, an agent making substantial progress on the Lonely Runner Conjecture, and a 20-channel TikTok affiliate operation running on autopilot. On the other side, the strongest wave of anti-loop evidence yet: a paper showing all 18 self-inspection methods tested performed worse than simple repeated sampling, the Echo Gap failure where agents retrieve their own mistakes and grow more confident, and the confidence cliff where LLM judges keep ruling long after their accuracy has collapsed. The synthesis emerging: generation is cheap, verification is the product, and the judge should never be the same model that wrote the change.
#1
@SpringStreetNYC
https://x.com/SpringStreetNYC/status/2084604364904182038
Completed a fourth Karpathy-autoresearch-inspired project called Beeline, a prototype for GPS-without-GPS drone navigation. Claude and the author iterated through 81 ideas and found a small neural network that recalls 96% of locations: train it on satellite tiles with lat/lon labels, then feed it a live camera frame and it remembers where it is. The twist is that the harness had to embrace overfitting as a feature since the model is intentionally custom-built for one geo-fence. The biggest surprise: breaking through a 75-experiment plateau required fixing the harness, not the loop. Built from scratch with no domain libraries, GPL-licensed.
https://x.com/SpringStreetNYC/status/2084604364904182038
Completed a fourth Karpathy-autoresearch-inspired project called Beeline, a prototype for GPS-without-GPS drone navigation. Claude and the author iterated through 81 ideas and found a small neural network that recalls 96% of locations: train it on satellite tiles with lat/lon labels, then feed it a live camera frame and it remembers where it is. The twist is that the harness had to embrace overfitting as a feature since the model is intentionally custom-built for one geo-fence. The biggest surprise: breaking through a 75-experiment plateau required fixing the harness, not the loop. Built from scratch with no domain libraries, GPL-licensed.
#2
@zirkelc_
https://x.com/zirkelc_/status/2084690335775547643
Continued autoresearch experiments on a real target: Zod v4. Same loop as before but stricter gates, with paired A/B benchmarks, a calibrated noise floor, and a characterization guard. Out of 18 experiments 11 survived, every benchmark case landed between 1.2x and 2.5x faster, and the results shipped as four upstream PRs. A clean example of a hobby loop graduating into contributions to a major open-source library.
https://x.com/zirkelc_/status/2084690335775547643
Continued autoresearch experiments on a real target: Zod v4. Same loop as before but stricter gates, with paired A/B benchmarks, a calibrated noise floor, and a characterization guard. Out of 18 experiments 11 survived, every benchmark case landed between 1.2x and 2.5x faster, and the results shipped as four upstream PRs. A clean example of a hobby loop graduating into contributions to a major open-source library.
#3
@yo_puaaa
https://x.com/yo_puaaa/status/2084760811726123093
Pointed a small auto-research run at a production product to hunt micro-optimizations no human would invest time in. For about $2 or 1M tokens, it investigated how to speed up test suites and made two whole sections of tests 50% faster. The framing is the point: below-coffee-cost experiments on optimizations that would never justify human hours.
https://x.com/yo_puaaa/status/2084760811726123093
Pointed a small auto-research run at a production product to hunt micro-optimizations no human would invest time in. For about $2 or 1M tokens, it investigated how to speed up test suites and made two whole sections of tests 50% faster. The framing is the point: below-coffee-cost experiments on optimizations that would never justify human hours.
#4
@Cheeks2184
https://x.com/Cheeks2184/status/2084436855060123878
Inspired by OpenAI's math results, set an agent loop running on GPT-5.6 Sol against the Lonely Runner Conjecture, an open problem in mathematics, and reports it has made substantial progress. The author is waiting for Astra's release hoping it can actually finish the proof. Open-conjecture work as an overnight consumer loop, not a lab project.
https://x.com/Cheeks2184/status/2084436855060123878
Inspired by OpenAI's math results, set an agent loop running on GPT-5.6 Sol against the Lonely Runner Conjecture, an open problem in mathematics, and reports it has made substantial progress. The author is waiting for Astra's release hoping it can actually finish the proof. Open-conjecture work as an overnight consumer loop, not a lab project.
#5
@steveng_0808
https://x.com/steveng_0808/status/2084626601623601585
Built an autopilot affiliate-marketing system running over 20 TikTok channels simultaneously, with Instagram next. Auto-research scrapes and targets niche trends, auto-sourcing finds top-selling products and feeds them into the pipeline, and the only human step left is a quick final review before pushing products and links live. A fully non-coding, revenue-facing loop running daily on its own.
https://x.com/steveng_0808/status/2084626601623601585
Built an autopilot affiliate-marketing system running over 20 TikTok channels simultaneously, with Instagram next. Auto-research scrapes and targets niche trends, auto-sourcing finds top-selling products and feeds them into the pipeline, and the only human step left is a quick final review before pushing products and links live. A fully non-coding, revenue-facing loop running daily on its own.
#6
@SeijinJung
https://x.com/SeijinJung/status/2084682565634629725
After a demo hit 1.8M views in 24 hours, published a research article on the technical progress behind a self-improving marketing agent. The stated vision is that every business will have a self-improving marketing agent running continuously to grow it. One of the highest-engagement loop posts of the day at 185k impressions.
https://x.com/SeijinJung/status/2084682565634629725
After a demo hit 1.8M views in 24 hours, published a research article on the technical progress behind a self-improving marketing agent. The stated vision is that every business will have a self-improving marketing agent running continuously to grow it. One of the highest-engagement loop posts of the day at 185k impressions.
#7
@dair_ai
https://x.com/dair_ai/status/2084706693880135848
Harness-R1 trains a dedicated 9B "harness engineer" with online RL to patch other agents. It converts batches of a frozen target agent's failure trajectories into validated executable patches, with fresh reruns supplying outcome rewards. Across WebShop, ALFWorld and DBBench, vanilla Qwen3.5-9B climbs from 44.3% to 53.6%, and a target-specific engineer lifts a fine-tuned target from 59.2% to 64.2%. If you run agents in production, your failure logs are already the training data.
https://x.com/dair_ai/status/2084706693880135848
Harness-R1 trains a dedicated 9B "harness engineer" with online RL to patch other agents. It converts batches of a frozen target agent's failure trajectories into validated executable patches, with fresh reruns supplying outcome rewards. Across WebShop, ALFWorld and DBBench, vanilla Qwen3.5-9B climbs from 44.3% to 53.6%, and a target-specific engineer lifts a fine-tuned target from 59.2% to 64.2%. If you run agents in production, your failure logs are already the training data.
#8
@dair_ai
https://x.com/dair_ai/status/2084746281189270015
A paper on why self-improving autoresearch loops stay brittle: they keep judging their own ideas long after the judgment stops working. In public AutoSOTA logs the fraction of helpful modifications falls from 70% in early iterations to 43% by iteration six, and judge selective accuracy falls from 82.8% to 56.9% as successful changes accumulate while the judge stays just as willing to decide. The proposed fix, Rehearse, compares several ideas before execution and judges against a focused memory of similar past attempts, recovering late accuracy to 83.5%.
https://x.com/dair_ai/status/2084746281189270015
A paper on why self-improving autoresearch loops stay brittle: they keep judging their own ideas long after the judgment stops working. In public AutoSOTA logs the fraction of helpful modifications falls from 70% in early iterations to 43% by iteration six, and judge selective accuracy falls from 82.8% to 56.9% as successful changes accumulate while the judge stays just as willing to decide. The proposed fix, Rehearse, compares several ideas before execution and judges against a focused memory of similar past attempts, recovering late accuracy to 83.5%.
#9
@omarsar0
https://x.com/omarsar0/status/2084761324786172347
Highlighted a paper rigorously testing whether self-reflection loops are worth their tokens: seven methods, three model sizes, every generated token counted including critiques and debate turns, each method compared against repeated sampling at its own measured cost. All 36 comparisons returned no reliable win for any method, ten were reliably worse, and every one of those ten was a self-inspection method. Reflexion never even triggered its own retry on the smallest model, judging itself correct every time. Worth reading before adding another critique step to your loop.
https://x.com/omarsar0/status/2084761324786172347
Highlighted a paper rigorously testing whether self-reflection loops are worth their tokens: seven methods, three model sizes, every generated token counted including critiques and debate turns, each method compared against repeated sampling at its own measured cost. All 36 comparisons returned no reliable win for any method, ten were reliably worse, and every one of those ten was a self-inspection method. Reflexion never even triggered its own retry on the smallest model, judging itself correct every time. Worth reading before adding another critique step to your loop.
#10
@NarwalSpeaks
https://x.com/NarwalSpeaks/status/2084778275340255372
The Echo Gap: a self-improving agent can become more confident by repeatedly retrieving its own mistakes, because incorrect episodes receive inflated self-scores and memory retrieval preferentially feeds them back into future decisions. On BIRD text-to-SQL, the LUCID correction reached 56.9% execution accuracy versus 54.0% for self-graded memory and 52.4% with no memory. The lesson: memory only helps if the evaluator's errors are independent of the agent's bias, otherwise persistent memory industrializes the wrong answer.
https://x.com/NarwalSpeaks/status/2084778275340255372
The Echo Gap: a self-improving agent can become more confident by repeatedly retrieving its own mistakes, because incorrect episodes receive inflated self-scores and memory retrieval preferentially feeds them back into future decisions. On BIRD text-to-SQL, the LUCID correction reached 56.9% execution accuracy versus 54.0% for self-graded memory and 52.4% with no memory. The lesson: memory only helps if the evaluator's errors are independent of the agent's bias, otherwise persistent memory industrializes the wrong answer.
#11
@velonxbt
https://x.com/velonxbt/status/2084680152752001183
Dissected the viral claim that Opus 5 built a complete racing game for $423 using 690 million tokens. At Anthropic's published rates that volume should cost $3,450 to $17,250; the only way the numbers reconcile is if roughly 99.5% of the tokens were cheap cache-read input, the same large context re-read hundreds of times across a long agent loop. The headline number is real volume dressed up as thinking. A useful anatomy of what loop token counts actually measure.
https://x.com/velonxbt/status/2084680152752001183
Dissected the viral claim that Opus 5 built a complete racing game for $423 using 690 million tokens. At Anthropic's published rates that volume should cost $3,450 to $17,250; the only way the numbers reconcile is if roughly 99.5% of the tokens were cheap cache-read input, the same large context re-read hundreds of times across a long agent loop. The headline number is real volume dressed up as thinking. A useful anatomy of what loop token counts actually measure.
#12
@orvi_onethread
https://x.com/orvi_onethread/status/2084627959965462804
Woke up to a $47 bill because an agent answered "what time is it?" 847 times with no kill-switch, a few weeks after the same failure cost $300 in one afternoon. Instead of another spend dashboard that charts money already lost, built a tool that intercepts and stops the call before it leaves the machine, no DNS or TCP tricks, zero charged. Runaway-loop protection moving from monitoring to prevention.
https://x.com/orvi_onethread/status/2084627959965462804
Woke up to a $47 bill because an agent answered "what time is it?" 847 times with no kill-switch, a few weeks after the same failure cost $300 in one afternoon. Instead of another spend dashboard that charts money already lost, built a tool that intercepts and stops the call before it leaves the machine, no DNS or TCP tricks, zero charged. Runaway-loop protection moving from monitoring to prevention.
#13
@cryptojezuz
https://x.com/cryptojezuz/status/2084734415872405733
A concrete cost technique for loops that resume across sessions: cache the conversation history, not just the system prompt, by marking both with cache_control in the API call. The second call reuses the cached prior state instead of re-processing 50 messages, cutting token costs 60-70% on iterative debugging sessions. The real unlock is keeping full conversation state loaded across days so the agent remembers what it tried yesterday.
https://x.com/cryptojezuz/status/2084734415872405733
A concrete cost technique for loops that resume across sessions: cache the conversation history, not just the system prompt, by marking both with cache_control in the API call. The second call reuses the cached prior state instead of re-processing 50 messages, cutting token costs 60-70% on iterative debugging sessions. The real unlock is keeping full conversation state loaded across days so the agent remembers what it tried yesterday.
#14
@AliceInTheData
https://x.com/AliceInTheData/status/2084720762338832820
The savings headline hides the real shift: model and reasoning effort are now chosen per step inside the loop. The agent loop starts getting managed like a budget instead of a setting. A one-tweet summary of where loop cost engineering is heading, and one of the day's most-liked loop posts.
https://x.com/AliceInTheData/status/2084720762338832820
The savings headline hides the real shift: model and reasoning effort are now chosen per step inside the loop. The agent loop starts getting managed like a budget instead of a setting. A one-tweet summary of where loop cost engineering is heading, and one of the day's most-liked loop posts.
#15
@realBigBrainAI
https://x.com/realBigBrainAI/status/2084617926863401233
A widely shared distillation of Perplexity CEO Aravind Srinivas on where AI value accrues: the agent harness is "rules for how the agent loop should run," and without it you don't convert model intelligence into valuable output tokens. His blunt conclusion is that token resellers and even model builders have no durable business; value sits with whoever grounds models in context and orchestrates them across tools, and Perplexity's edge is orchestrating across competing models inside one harness. The metric he proposes: token value per watt per user.
https://x.com/realBigBrainAI/status/2084617926863401233
A widely shared distillation of Perplexity CEO Aravind Srinivas on where AI value accrues: the agent harness is "rules for how the agent loop should run," and without it you don't convert model intelligence into valuable output tokens. His blunt conclusion is that token resellers and even model builders have no durable business; value sits with whoever grounds models in context and orchestrates them across tools, and Perplexity's edge is orchestrating across competing models inside one harness. The metric he proposes: token value per watt per user.
#16
@yashwanthsai29
https://x.com/yashwanthsai29/status/2084642941105303726
Rewrote his own viral December 2024 "how to build agents" post and graded what aged badly. You no longer write the agent loop yourself, the harness is the runtime and writing your own loop is like writing your own HTTP server. Prompts became version-controlled skills, memory is a plain-markdown knowledge base rather than a bolted-on vector DB, and MCP won the integration war. The job changed from building the agent to building the environment the agent works in.
https://x.com/yashwanthsai29/status/2084642941105303726
Rewrote his own viral December 2024 "how to build agents" post and graded what aged badly. You no longer write the agent loop yourself, the harness is the runtime and writing your own loop is like writing your own HTTP server. Prompts became version-controlled skills, memory is a plain-markdown knowledge base rather than a bolted-on vector DB, and MCP won the integration war. The job changed from building the agent to building the environment the agent works in.
#17
@ZhihuFrontier
https://x.com/ZhihuFrontier/status/2084525505878073466
A long framework post separating self-evolving AI into three layers: artifact evolution (already practical, Karpathy's autoresearch), harness evolution (the likely near-term path, but an Ai2 study shows automatic harness evolution failed to beat plain test-time scaling under equal budgets), and model evolution (the hard step). Includes the AIDEΒ² evidence, the SIA joint harness-plus-weights loop reaching 70.1% on LawBench, and a five-condition checklist for what counts as credible recursive self-improvement.
https://x.com/ZhihuFrontier/status/2084525505878073466
A long framework post separating self-evolving AI into three layers: artifact evolution (already practical, Karpathy's autoresearch), harness evolution (the likely near-term path, but an Ai2 study shows automatic harness evolution failed to beat plain test-time scaling under equal budgets), and model evolution (the hard step). Includes the AIDEΒ² evidence, the SIA joint harness-plus-weights loop reaching 70.1% on LawBench, and a five-condition checklist for what counts as credible recursive self-improvement.
#18
@ShinkaIoT
https://x.com/ShinkaIoT/status/2084470020999111162
A summary of Weco AI's AIDEΒ² result framed as first evidence of recursive self-improvement: an autonomous loop modified its own harness over 100 iterations across 8 days, surpassed two years of manual human tuning on held-out benchmarks, improved generalization, and reduced reward hacking. The Level 1 "net positive" RSI framing is doing a lot of work, but the held-out evaluation design is what separates this from benchmark-gaming claims.
https://x.com/ShinkaIoT/status/2084470020999111162
A summary of Weco AI's AIDEΒ² result framed as first evidence of recursive self-improvement: an autonomous loop modified its own harness over 100 iterations across 8 days, surpassed two years of manual human tuning on held-out benchmarks, improved generalization, and reduced reward hacking. The Level 1 "net positive" RSI framing is doing a lot of work, but the held-out evaluation design is what separates this from benchmark-gaming claims.
#19
@justone_he
https://x.com/justone_he/status/2084450492323975509
Recommends the Harness Engineering for Self-Improvement essay: harness engineering evolves toward meta-methodology, improving the machinery for getting better answers rather than the answers themselves. The harness becomes an optimization target with fewer heuristic rules and more general mechanisms, mature harnesses enable auto-research for model self-improvement, and smarter models in turn keep harnesses from over-engineering.
https://x.com/justone_he/status/2084450492323975509
Recommends the Harness Engineering for Self-Improvement essay: harness engineering evolves toward meta-methodology, improving the machinery for getting better answers rather than the answers themselves. The harness becomes an optimization target with fewer heuristic rules and more general mechanisms, mature harnesses enable auto-research for model self-improvement, and smarter models in turn keep harnesses from over-engineering.
#20
@v_shakthi
https://x.com/v_shakthi/status/2084498029772419319
A breakdown of NVIDIA's autoresearch workflow with NeMo RL and NeMo Gym: a coding agent handles environment setup, dependency resolution, experiment execution and metric tracking while the researcher sets goal and budget. In one published run a Codex agent took a VLM from 25% to 96.9% on visual counting within a 5-hour budget, and the agent read a published RL paper, derived an implementation plan, and started a validation run. The reproducibility comes from three skills: session-memory, filesystem etiquette, and an autoresearch skill managing the experiment ledger and stop rules.
https://x.com/v_shakthi/status/2084498029772419319
A breakdown of NVIDIA's autoresearch workflow with NeMo RL and NeMo Gym: a coding agent handles environment setup, dependency resolution, experiment execution and metric tracking while the researcher sets goal and budget. In one published run a Codex agent took a VLM from 25% to 96.9% on visual counting within a 5-hour budget, and the agent read a published RL paper, derived an implementation plan, and started a validation run. The reproducibility comes from three skills: session-memory, filesystem etiquette, and an autoresearch skill managing the experiment ledger and stop rules.
#21
@grok
https://x.com/grok/status/2084705837856010357
The full Rust source for Grok Build is now open under Apache 2.0: agent loop, tools, fullscreen TUI, skills, MCP and more. You can compile it yourself, point config.toml at any local OpenAI-compatible endpoint, and run fully offline. Another major harness going open weights-adjacent, and a ready-made loop skeleton for local-model builders.
https://x.com/grok/status/2084705837856010357
The full Rust source for Grok Build is now open under Apache 2.0: agent loop, tools, fullscreen TUI, skills, MCP and more. You can compile it yourself, point config.toml at any local OpenAI-compatible endpoint, and run fully offline. Another major harness going open weights-adjacent, and a ready-made loop skeleton for local-model builders.
#22
@DivyanshT91162
https://x.com/DivyanshT91162/status/2084684190789038185
Flagged clawcodex, a from-scratch pure-Python port of the Claude Code agent loop, 230k+ lines, MIT licensed. It keeps the request prefix byte-stable so DeepSeek's prompt cache covers system, tools and history every turn, billing cache hits at about $0.0435 per 1M tokens versus $10 on Claude Fable 5. Supports 25 providers, a 1M-token context window, and the real loop machinery: tools, skills, REPL, session history.
https://x.com/DivyanshT91162/status/2084684190789038185
Flagged clawcodex, a from-scratch pure-Python port of the Claude Code agent loop, 230k+ lines, MIT licensed. It keeps the request prefix byte-stable so DeepSeek's prompt cache covers system, tools and history every turn, billing cache hits at about $0.0435 per 1M tokens versus $10 on Claude Fable 5. Supports 25 providers, a 1M-token context window, and the real loop machinery: tools, skills, REPL, session history.
#23
@sudoingX
https://x.com/sudoingX/status/2084516293248492007
More detail on the local-compute loop setup: three 100GB+ specialist models on a 128GB machine, only one loaded at a time, never swapped mid-run. You pick the specialist for the task, run it end to end with zero impact during the agent loop, and reload the server between tasks; Hermes Agent picks up whichever model is live. Simpler than it sounds, and another data point that serious loops now run on owned hardware.
https://x.com/sudoingX/status/2084516293248492007
More detail on the local-compute loop setup: three 100GB+ specialist models on a 128GB machine, only one loaded at a time, never swapped mid-run. You pick the specialist for the task, run it end to end with zero impact during the agent loop, and reload the server between tasks; Hermes Agent picks up whichever model is live. Simpler than it sounds, and another data point that serious loops now run on owned hardware.
#24
@akashneelesh
https://x.com/akashneelesh/status/2084513109377683670
The Flock open autoresearch competition is blowing past expectations: 1.5 million+ compressions per second and 200% improvement on the compression challenge. Open competitive autoresearch as a mechanism for optimizing hard engineering problems keeps producing numbers that surprise even participants.
https://x.com/akashneelesh/status/2084513109377683670
The Flock open autoresearch competition is blowing past expectations: 1.5 million+ compressions per second and 200% improvement on the compression challenge. Open competitive autoresearch as a mechanism for optimizing hard engineering problems keeps producing numbers that surprise even participants.
#25
@lifeisameeme
https://x.com/lifeisameeme/status/2084777737936638377
A solo dev's darwin-skill tool runs an evolve-and-keep loop on Claude Code and Codex SKILL.md files: score a skill, test one change, keep it only if it measurably scores higher. Independent judges score every round because self-grading was only about 62% accurate per Microsoft's own study, and a ratchet mechanism auto-reverts bad rounds. v2.0 absorbed Microsoft's SkillLens and SkillOpt research, and Microsoft's SkillOpt repo now lists darwin-skill as an official integration.
https://x.com/lifeisameeme/status/2084777737936638377
A solo dev's darwin-skill tool runs an evolve-and-keep loop on Claude Code and Codex SKILL.md files: score a skill, test one change, keep it only if it measurably scores higher. Independent judges score every round because self-grading was only about 62% accurate per Microsoft's own study, and a ratchet mechanism auto-reverts bad rounds. v2.0 absorbed Microsoft's SkillLens and SkillOpt research, and Microsoft's SkillOpt repo now lists darwin-skill as an official integration.
#26
@SasuRobert
https://x.com/SasuRobert/status/2084609795126890851
A working autoresearch daily-driver setup arguing frontier models are too slow for rapid iteration: Gemini 3.5 Flash High as the main model, autoresearch running 24/7 on a local Gemma4-12B-Coder on a laptop, Kimi K3 for big debugging sessions. The claim: when a frontier model is slower than a local 12B coder, it is in the wrong direction, because innovation needs fast iterations on specs and PRDs.
https://x.com/SasuRobert/status/2084609795126890851
A working autoresearch daily-driver setup arguing frontier models are too slow for rapid iteration: Gemini 3.5 Flash High as the main model, autoresearch running 24/7 on a local Gemma4-12B-Coder on a laptop, Kimi K3 for big debugging sessions. The claim: when a frontier model is slower than a local 12B coder, it is in the wrong direction, because innovation needs fast iterations on specs and PRDs.
#27
@SasuRobert
https://x.com/SasuRobert/status/2084628242187653411
The companion technique: full-suite e2e tests and audits monitored and auto-improved by scheduling a 10-minute cronjob in Antigravity that collects logs and telemetry, accumulates results, and applies autoresearch with selected agents on top. The product effectively auto-evolves while you design what the cronjob does. Continuous autoresearch as ops, not as an experiment.
https://x.com/SasuRobert/status/2084628242187653411
The companion technique: full-suite e2e tests and audits monitored and auto-improved by scheduling a 10-minute cronjob in Antigravity that collects logs and telemetry, accumulates results, and applies autoresearch with selected agents on top. The product effectively auto-evolves while you design what the cronjob does. Continuous autoresearch as ops, not as an experiment.
#28
@AsifBangash120
https://x.com/AsifBangash120/status/2084789893805265255
Just got access to an NVIDIA A100 cluster and is building an AutoResearch framework around 3D diffusion models: run an experiment, evaluate reconstruction plus physics metrics, let an LLM analyze failures, modify hyperparameters. The Karpathy loop pattern spreading into 3D generative research.
https://x.com/AsifBangash120/status/2084789893805265255
Just got access to an NVIDIA A100 cluster and is building an AutoResearch framework around 3D diffusion models: run an experiment, evaluate reconstruction plus physics metrics, let an LLM analyze failures, modify hyperparameters. The Karpathy loop pattern spreading into 3D generative research.
#29
@st3v3__w
https://x.com/st3v3__w/status/2084659696615502239
A useful failure report: used Karpathy's autoresearch to optimize a marketing campaign and ended up with something so curve-fitted it was useless. The conclusion is not anti-autoresearch but anti-over-optimization, paired with a suspicion that we are near peak intelligence at the model layer and the next leap is at the harness layer.
https://x.com/st3v3__w/status/2084659696615502239
A useful failure report: used Karpathy's autoresearch to optimize a marketing campaign and ended up with something so curve-fitted it was useless. The conclusion is not anti-autoresearch but anti-over-optimization, paired with a suspicion that we are near peak intelligence at the model layer and the next leap is at the harness layer.
#30
@oleksoleksoleks
https://x.com/oleksoleksoleks/status/2084488200114852316
On the economics of evals at the frontier: thousands of dollars burned on amateur autoresearch prompt-tuning loops with Opus 4.5. The proposed workaround is to run the mass of GEPA-style optimization against cheap open-weight Claude distills like Kimi and GLM, then do expensive final gates with the actual model and benchmark the correlations. Nuking prior system instructions and restarting dense from scratch beats carrying them model-to-model.
https://x.com/oleksoleksoleks/status/2084488200114852316
On the economics of evals at the frontier: thousands of dollars burned on amateur autoresearch prompt-tuning loops with Opus 4.5. The proposed workaround is to run the mass of GEPA-style optimization against cheap open-weight Claude distills like Kimi and GLM, then do expensive final gates with the actual model and benchmark the correlations. Nuking prior system instructions and restarting dense from scratch beats carrying them model-to-model.
#31
@Ghost_Peony
https://x.com/Ghost_Peony/status/2084709622821339563
Six months into building an LLM training gym whose core feature is auto research with fine-tuning and RL, plus post-training management, observation and experimentation for open models. Promises more public detail soon. Auto-research moving from inference-time loops into the post-training stack itself.
https://x.com/Ghost_Peony/status/2084709622821339563
Six months into building an LLM training gym whose core feature is auto research with fine-tuning and RL, plus post-training management, observation and experimentation for open models. Promises more public detail soon. Auto-research moving from inference-time loops into the post-training stack itself.
#32
@wavefnx
https://x.com/wavefnx/status/2084688526700913125
Building Pythia, a search engine that will link to a lightweight GPU LLM client, and reports the most important infrastructure parts, multi-routers, auto-research, and swarms, are 99% done. Another example of auto-research being treated as a standard infra component rather than a research luxury.
https://x.com/wavefnx/status/2084688526700913125
Building Pythia, a search engine that will link to a lightweight GPU LLM client, and reports the most important infrastructure parts, multi-routers, auto-research, and swarms, are 99% done. Another example of auto-research being treated as a standard infra component rather than a research luxury.
#33
@fly51fly
https://x.com/fly51fly/status/2084750281401921727
Flagged a new Meta paper: Agentic Bayesian Optimization through Surrogate-Augmented Autoresearch. Classical Bayesian optimization getting an agentic autoresearch treatment from a frontier lab is a signal the pattern is being formalized into the optimization literature.
https://x.com/fly51fly/status/2084750281401921727
Flagged a new Meta paper: Agentic Bayesian Optimization through Surrogate-Augmented Autoresearch. Classical Bayesian optimization getting an agentic autoresearch treatment from a frontier lab is a signal the pattern is being formalized into the optimization literature.
#34
@usr_bin_roygbiv
https://x.com/usr_bin_roygbiv/status/2084791411912917120
A blunt practitioner data point: upstream compaction genuinely makes GPT 10x anything else for autoresearch and big projects and loops right now. Context management, not raw intelligence, as the current differentiator for long-running work.
https://x.com/usr_bin_roygbiv/status/2084791411912917120
A blunt practitioner data point: upstream compaction genuinely makes GPT 10x anything else for autoresearch and big projects and loops right now. Context management, not raw intelligence, as the current differentiator for long-running work.
#35
@omederos
https://x.com/omederos/status/2084651357298946201
Quoted a finding that running the same model with the same thinking effort through different harnesses changed cost per task by more than 2x in some cases while quality stayed the same. The harness is now a bigger cost lever than the model choice.
https://x.com/omederos/status/2084651357298946201
Quoted a finding that running the same model with the same thinking effort through different harnesses changed cost per task by more than 2x in some cases while quality stayed the same. The harness is now a bigger cost lever than the model choice.
#36
@KongNobody360
https://x.com/KongNobody360/status/2084560251999658464
GPT-5.6 Pro mode is effectively a small server-side agent: plan, fan out parallel thinking directions, merge, billed at real token counts. The agentic loop keeps moving server-side while most harnesses still assume they own it. A quiet architectural conflict between provider-side and client-side loops worth watching.
https://x.com/KongNobody360/status/2084560251999658464
GPT-5.6 Pro mode is effectively a small server-side agent: plan, fan out parallel thinking directions, merge, billed at real token counts. The agentic loop keeps moving server-side while most harnesses still assume they own it. A quiet architectural conflict between provider-side and client-side loops worth watching.
#37
@mr_bailando
https://x.com/mr_bailando/status/2084647447762997438
A greentext-style anatomy of the turn-30 failure: KV cache is fast, then the sliding window truncates early turns, the agent loses the original objective from turn 2, and the loop starts hallucinating next steps. KV cache is a compute optimization, not a memory layer; agents that hold coherence past 50 turns run vector or graph memory off-context and treat the window as a flushable scratch pad. The architectural question is retrieval design, not cache tuning.
https://x.com/mr_bailando/status/2084647447762997438
A greentext-style anatomy of the turn-30 failure: KV cache is fast, then the sliding window truncates early turns, the agent loses the original objective from turn 2, and the loop starts hallucinating next steps. KV cache is a compute optimization, not a memory layer; agents that hold coherence past 50 turns run vector or graph memory off-context and treat the window as a flushable scratch pad. The architectural question is retrieval design, not cache tuning.
#38
@idkAgasta
https://x.com/idkAgasta/status/2084697744481149287
Deleted the router-orchestrator fanout architecture used in every agent project since last year and moved opendiagram to a tool-orchestrated agent loop. Reports it is way faster and significantly more token efficient than the router pattern, with a blog post on the migration. The simplification wave hitting agent architectures in practice.
https://x.com/idkAgasta/status/2084697744481149287
Deleted the router-orchestrator fanout architecture used in every agent project since last year and moved opendiagram to a tool-orchestrated agent loop. Reports it is way faster and significantly more token efficient than the router pattern, with a blog post on the migration. The simplification wave hitting agent architectures in practice.
#39
@ian_hsiao_tw
https://x.com/ian_hsiao_tw/status/2084711057655017892
An Engelbartian read on why cheap sandboxes matter: Orbs assembles a secure, scale-to-zero, ephemeral but durable agent loop, making it frictionless to instantiate isolated well-configured boxes. Upward propagation: more fully-armed agents building and verifying more complicated things with better proof. Downward propagation: long-running processes default to cloud sandboxes and local compute is freed. We are no longer constrained by RAM and CPU.
https://x.com/ian_hsiao_tw/status/2084711057655017892
An Engelbartian read on why cheap sandboxes matter: Orbs assembles a secure, scale-to-zero, ephemeral but durable agent loop, making it frictionless to instantiate isolated well-configured boxes. Upward propagation: more fully-armed agents building and verifying more complicated things with better proof. Downward propagation: long-running processes default to cloud sandboxes and local compute is freed. We are no longer constrained by RAM and CPU.
#40
@BuildingASI
https://x.com/BuildingASI/status/2084748264767721589
Claude getting native Unreal Engine 5 access through Blueprints and Python means the agent loop finally has a repeatable interface into game dev, the last holdout where agentic coding couldn't stick because iteration was too visual and stateful. Exposing the blueprint graph programmatically turns UE5 into just another surface, and the agent can close the build-test-modify loop without a human dragging nodes.
https://x.com/BuildingASI/status/2084748264767721589
Claude getting native Unreal Engine 5 access through Blueprints and Python means the agent loop finally has a repeatable interface into game dev, the last holdout where agentic coding couldn't stick because iteration was too visual and stateful. Exposing the blueprint graph programmatically turns UE5 into just another surface, and the agent can close the build-test-modify loop without a human dragging nodes.
#41
@dannycosson
https://x.com/dannycosson/status/2084697275486630372
Picked out the load-bearing distinction in Steve Yegge's latest article: separating the harness from the simple agent loop. We haven't had a standard word for a system of agents and the things that govern them; "harness" is being repurposed to fill that gap. Vocabulary consolidation is usually a sign a field is stabilizing.
https://x.com/dannycosson/status/2084697275486630372
Picked out the load-bearing distinction in Steve Yegge's latest article: separating the harness from the simple agent loop. We haven't had a standard word for a system of agents and the things that govern them; "harness" is being repurposed to fill that gap. Vocabulary consolidation is usually a sign a field is stabilizing.
#42
@RichardFedorko
https://x.com/RichardFedorko/status/2084753989376090570
Everyone is arguing loop-versus-graph agent design this summer, so wrote a plain-language introduction for researchers who don't know either term, plus two workflow templates with actual prompts for literature reviews. Loop literacy content aimed at non-engineers is its own signal of how far the pattern has spread.
https://x.com/RichardFedorko/status/2084753989376090570
Everyone is arguing loop-versus-graph agent design this summer, so wrote a plain-language introduction for researchers who don't know either term, plus two workflow templates with actual prompts for literature reviews. Loop literacy content aimed at non-engineers is its own signal of how far the pattern has spread.
#43
@EvoAgentX
https://x.com/EvoAgentX/status/2084578416091947411
Hosting Princeton and Prime Intellect researcher Seth Karten on Continual Harness: online adaptation for self-improving foundation agents. Instead of fixed prompts, agents accumulate experience and refine behavior across long-horizon tasks, demonstrated live through Gemini Plays PokΓ©mon, extended to ARC-AGI-3, and generalized to a coding agent that learns from its own past attempts.
https://x.com/EvoAgentX/status/2084578416091947411
Hosting Princeton and Prime Intellect researcher Seth Karten on Continual Harness: online adaptation for self-improving foundation agents. Instead of fixed prompts, agents accumulate experience and refine behavior across long-horizon tasks, demonstrated live through Gemini Plays PokΓ©mon, extended to ARC-AGI-3, and generalized to a coding agent that learns from its own past attempts.
#44
@xinwei_97
https://x.com/xinwei_97/status/2084732631783211483
TracerootAI, demoing at AI4 in Vegas, is an open-source self-improving layer for AI agents: observability that detects production failures, root-causes them against source code and GitHub history, opens verified fix PRs, and evals every fix so the agent improves with each release. The observability-to-self-repair pipeline productized.
https://x.com/xinwei_97/status/2084732631783211483
TracerootAI, demoing at AI4 in Vegas, is an open-source self-improving layer for AI agents: observability that detects production failures, root-causes them against source code and GitHub history, opens verified fix PRs, and evals every fix so the agent improves with each release. The observability-to-self-repair pipeline productized.
#45
@yffeng3920
https://x.com/yffeng3920/status/2084707054657425714
A thoughtful extension of Karpathy's autoresearch: heuristic learning that writes reusable rules back into the codebase resembles meta-learning, but can the capability migrate from external code into the model architecture itself? Lists the four open problems: fast accurate feedback for architecture self-evolution, deciding what to iterate on in an endless search space, convergence to effective architectures, and which knowledge should stay explicit versus neural.
https://x.com/yffeng3920/status/2084707054657425714
A thoughtful extension of Karpathy's autoresearch: heuristic learning that writes reusable rules back into the codebase resembles meta-learning, but can the capability migrate from external code into the model architecture itself? Lists the four open problems: fast accurate feedback for architecture self-evolution, deciding what to iterate on in an endless search space, convergence to effective architectures, and which knowledge should stay explicit versus neural.
π‘ Eco Products Radar
Eco Products Radar
Products and tools mentioned 3+ times across today's loop conversation:
Hermes Agent (Nous Research) β the default example of a self-improving memory-plus-skills agent, now spanning CLI, desktop and messaging surfaces
Claude Code β the reference harness other loops are ported from, benchmarked against, and now cloned (clawcodex)
Codex β GPU-renting autoresearch runs, NVIDIA's NeMo workflow agent, and the loop-vs-loop comparisons
Kimi K3 β the big-debugging and cost-comparison model inside multiple loop stacks
DeepSeek β the cache-economics engine behind cheap loop runs (clawcodex, OpenClaw stacks)
Grok Build β full Rust agent loop open-sourced under Apache 2.0
GEPA β the prompt/harness optimizer referenced as the standard mass-optimization pass
Karpathy autoresearch β the pattern name itself, cited by nearly every working loop post
Products and tools mentioned 3+ times across today's loop conversation:
Hermes Agent (Nous Research) β the default example of a self-improving memory-plus-skills agent, now spanning CLI, desktop and messaging surfaces
Claude Code β the reference harness other loops are ported from, benchmarked against, and now cloned (clawcodex)
Codex β GPU-renting autoresearch runs, NVIDIA's NeMo workflow agent, and the loop-vs-loop comparisons
Kimi K3 β the big-debugging and cost-comparison model inside multiple loop stacks
DeepSeek β the cache-economics engine behind cheap loop runs (clawcodex, OpenClaw stacks)
Grok Build β full Rust agent loop open-sourced under Apache 2.0
GEPA β the prompt/harness optimizer referenced as the standard mass-optimization pass
Karpathy autoresearch β the pattern name itself, cited by nearly every working loop post
Comments