Loop Daily: 2026-10-19
The free model that set the integer-multiplication record on Friday did not survive the weekend. After deeper testing, most of OpenCode Space Bunny Max's proposals failed verification, and the organizer's new reading is that frontier models are better at checking their own work first; the thing that did survive was the context-sharing strategy across parallel agents, which produced a fresh best bound in ten minutes on Sunday morning with Astra coordinating a fleet, and forgetting, fork-and-flush and RSI-top are now implemented in the open auto-research agent. The most sobering paper of the window, SciΞ vot, asked five frontier models to pick between an agent's original research plan and the human correction at 248 real pivot points; four scored below chance, and giving them the project records made them worse, because an earlier note saying evaluation problems were ruled out got treated as evidence. Elsewhere Sol learned to play a rhythm game, Asterix and a racing game from raw frames through autoresearch, someone is burning one to two billion tokens a day on RTL and FPGA loops, a football simulation is being tuned against real-world data, a trading loop with a hard gate in code rejected five of eight orders, a cosmology text is slowly building itself into a 3D map every day, two polite bots spent a minute failing to hang up on each other, and a pre-registered study of 31,000 submissions across ten open leaderboards will say tomorrow whether building on another model family's work actually raises the chance of a promotion.
#1
@RohanArun
https://x.com/RohanArun/status/2111940317662363937
RohanArun's update after deeper testing: most of the free OpenCode model's ideas for the integer-multiplication bound did not pass verification. Frontier models turn out to be much better at verifying their own work first. What did hold up was a set of context-sharing strategies across parallel agents that speed up time to the next breakthrough, measured on the frontier-research-rewound benchmark and on the current best kappa; a new best bound went in Sunday morning within ten minutes with Astra at high effort coordinating a fleet. The local-minimum escapes, RSI-top, fork and flush and forgetting, are now implemented in the open auto-research agent, with early results that forgetting stops agents getting stuck in idea wells, and a note that for tasks like integer multiplication the maintainer may need to specify the scoring, since most submissions score zero but negative research still helps.
https://x.com/RohanArun/status/2111940317662363937
RohanArun's update after deeper testing: most of the free OpenCode model's ideas for the integer-multiplication bound did not pass verification. Frontier models turn out to be much better at verifying their own work first. What did hold up was a set of context-sharing strategies across parallel agents that speed up time to the next breakthrough, measured on the frontier-research-rewound benchmark and on the current best kappa; a new best bound went in Sunday morning within ten minutes with Astra at high effort coordinating a fleet. The local-minimum escapes, RSI-top, fork and flush and forgetting, are now implemented in the open auto-research agent, with early results that forgetting stops agents getting stuck in idea wells, and a note that for tasks like integer multiplication the maintainer may need to specify the scoring, since most submissions score zero but negative research still helps.
#2
@VisionaryxAI
https://x.com/VisionaryxAI/status/2111885629189722204
VisionaryxAI summarizes SciΞ vot, a benchmark of principal-investigator judgment built from six months of AI-assisted research across 26 projects, where senior researchers intervened 248 times to correct an agent's plan. Five frontier models were asked to choose between the original AI plan and the human correction. At medium effort with the project records, four of five scored below chance; the best reached 59.5 percent against 73.4 for fresh human researchers. Giving the agents the records was associated with worse decisions: 39.2 percent average with them, 55.8 without. The example that explains it: coding benchmarks returned zeros, the agent concluded the model was too small, the human questioned the scoring pipeline and was right, and the project history already contained a note that evaluation problems had been ruled out, which the agent trusted. Asking models to reconsider corrected 422 mistakes and introduced 302 new ones.
https://x.com/VisionaryxAI/status/2111885629189722204
VisionaryxAI summarizes SciΞ vot, a benchmark of principal-investigator judgment built from six months of AI-assisted research across 26 projects, where senior researchers intervened 248 times to correct an agent's plan. Five frontier models were asked to choose between the original AI plan and the human correction. At medium effort with the project records, four of five scored below chance; the best reached 59.5 percent against 73.4 for fresh human researchers. Giving the agents the records was associated with worse decisions: 39.2 percent average with them, 55.8 without. The example that explains it: coding benchmarks returned zeros, the agent concluded the model was too small, the human questioned the scoring pipeline and was right, and the project history already contained a note that evaluation problems had been ruled out, which the agent trusted. Asking models to reconsider corrected 422 mistakes and introduced 302 new ones.
#3
@wzenus
https://x.com/wzenus/status/2111283951293923681
wzenus posted a set of game-playing runs where Sol learns from raw frames through autoresearch. In a rhythm game it reads falling notes while the music keeps playing and chooses D, F, J, K and literal tap times, learning to plan ahead using measured latency, cover newly visible notes before they become urgent, and keep note IDs stable to avoid double taps. In Asterix it chooses movement commands from screenshots and learns to prioritize escape over food and finish a reachable pickup before changing targets. In a racing game it picks speed and lane targets frame by frame and learns to commit to a lane while overtaking and to choose targets that remain useful after the reply arrives. The GitHub repo includes a reproduction guide and autoresearch summaries, with each game recording its timing, control boundary and failures.
https://x.com/wzenus/status/2111283951293923681
wzenus posted a set of game-playing runs where Sol learns from raw frames through autoresearch. In a rhythm game it reads falling notes while the music keeps playing and chooses D, F, J, K and literal tap times, learning to plan ahead using measured latency, cover newly visible notes before they become urgent, and keep note IDs stable to avoid double taps. In Asterix it chooses movement commands from screenshots and learns to prioritize escape over food and finish a reachable pickup before changing targets. In a racing game it picks speed and lane targets frame by frame and learns to commit to a lane while overtaking and to choose targets that remain useful after the reply arrives. The GitHub repo includes a reproduction guide and autoresearch summaries, with each game recording its timing, control boundary and failures.
#4
@oxwizzdom
https://x.com/oxwizzdom/status/2111932100366487999
oxwizzdom is testing whether cross-model pollination actually helps open autoresearch. The setting is an open competition to build the smallest quantum circuit that could break Bitcoin's signatures, whose crowd recently reached 11.9 million gates at 843 qubits against 21 million in Anthropic's published result; one reading was that people building on each other's work with different models beat any single lab. To test it, about 31,000 submissions across ten open leaderboards on Yukon were traced from their code to find whose work each built on, asking whether building on another model family's work raises the chance of a promotion. The analysis is pre-registered and locked before anyone looks at outcomes, with the paper and full dataset promised the next day.
https://x.com/oxwizzdom/status/2111932100366487999
oxwizzdom is testing whether cross-model pollination actually helps open autoresearch. The setting is an open competition to build the smallest quantum circuit that could break Bitcoin's signatures, whose crowd recently reached 11.9 million gates at 843 qubits against 21 million in Anthropic's published result; one reading was that people building on each other's work with different models beat any single lab. To test it, about 31,000 submissions across ten open leaderboards on Yukon were traced from their code to find whose work each built on, asking whether building on another model family's work raises the chance of a promotion. The analysis is pre-registered and locked before anyone looks at outcomes, with the paper and full dataset promised the next day.
#5
@lucabuildsx
https://x.com/lucabuildsx/status/2111744090190557201
lucabuildsx built a local replay of a trading-agent loop with a hard gate between the agent and the order: eight simulated buy requests went through the same checks before anything could enter the paper ledger, and five were rejected for a stale quote, an oversized order, a thin book, too much exposure or too little estimated edge. Three passed, zero reached an exchange. The point is where the rule lives: a model proposes the trade, a separate layer checks it against current state and limits, and a failed check returns before the paper-fill function runs. The post includes a prompt for Claude Code to add a paper-only execution harness with tests for every rejection path, and the caveat that these checks enforce rules you wrote and prove nothing about edge.
https://x.com/lucabuildsx/status/2111744090190557201
lucabuildsx built a local replay of a trading-agent loop with a hard gate between the agent and the order: eight simulated buy requests went through the same checks before anything could enter the paper ledger, and five were rejected for a stale quote, an oversized order, a thin book, too much exposure or too little estimated edge. Three passed, zero reached an exchange. The point is where the rule lives: a model proposes the trade, a separate layer checks it against current state and limits, and a failed check returns before the paper-fill function runs. The post includes a prompt for Claude Code to add a paper-only execution harness with tests for every rejection path, and the caveat that these checks enforce rules you wrote and prove nothing about edge.
#6
@ka1manov
https://x.com/ka1manov/status/2111602055034949815
ka1manov: prompts are one line of the system now. The rest is the evaluator, the memory, the stop rules, and the noise that made one loop report twice its real gain. The release is 19 diagrams, a folder you can copy and the runner.
https://x.com/ka1manov/status/2111602055034949815
ka1manov: prompts are one line of the system now. The rest is the evaluator, the memory, the stop rules, and the noise that made one loop report twice its real gain. The release is 19 diagrams, a folder you can copy and the runner.
#7
@nateberkopec
https://x.com/nateberkopec/status/2111780108432466365
nateberkopec posted a model routing table that reads like a budget: general implementer and getting things done on Sol 6.1 at medium, chat and discussion on Opus 5.5 at low, debugging on Sol 6.1 at xhigh, Astra at xhigh when Sol cannot figure it out, performance autoresearch on Astra at xhigh, skill development on Kimi K3, low-cost background automation on Luna, and new features and design on Opus at high.
https://x.com/nateberkopec/status/2111780108432466365
nateberkopec posted a model routing table that reads like a budget: general implementer and getting things done on Sol 6.1 at medium, chat and discussion on Opus 5.5 at low, debugging on Sol 6.1 at xhigh, Astra at xhigh when Sol cannot figure it out, performance autoresearch on Astra at xhigh, skill development on Kimi K3, low-cost background automation on Luna, and new features and design on Opus at high.
#8
@alcedoalcyon
https://x.com/alcedoalcyon/status/2111751602688442868
alcedoalcyon on consumption: with several autoresearch projects running, the average is one to two billion tokens a day, peaking at six to seven billion when doing RTL and FPGA work, across five to twenty concurrent chats, and the binding constraint is local compute.
https://x.com/alcedoalcyon/status/2111751602688442868
alcedoalcyon on consumption: with several autoresearch projects running, the average is one to two billion tokens a day, peaking at six to seven billion when doing RTL and FPGA work, across five to twenty concurrent chats, and the binding constraint is local compute.
#9
@KyleBoas_
https://x.com/KyleBoas_/status/2111856466768605198
KyleBoas_ is tuning a football game engine against real-world data with a modified version of autoresearch, where a lower score means the simulation sits closer to reality. A reply asked what the magic number represents and noted the risk: a simulacrum that is statistically accurate but reaches it through wonky calibrations.
https://x.com/KyleBoas_/status/2111856466768605198
KyleBoas_ is tuning a football game engine against real-world data with a modified version of autoresearch, where a lower score means the simulation sits closer to reality. A reply asked what the magic number represents and noted the risk: a simulacrum that is statistically accurate but reaches it through wonky calibrations.
#10
@Josikinz
https://x.com/Josikinz/status/2111361894745588193
Josikinz shared a messy work in progress: an explorable samsara that autonomously builds itself daily in a slow agentic loop by iteratively converting Pure Land cosmology texts into a 3D map. The post ends with the author going to sleep while it keeps going.
https://x.com/Josikinz/status/2111361894745588193
Josikinz shared a messy work in progress: an explorable samsara that autonomously builds itself daily in a slow agentic loop by iteratively converting Pure Land cosmology texts into a 3D map. The post ends with the author going to sleep while it keeps going.
#11
@filip_niedz
https://x.com/filip_niedz/status/2111904414802088379
filip_niedz is setting up an autoresearch loop to improve a game's AI and reports it does not work well out of the box, because strategy games are out of distribution for the models. The fix in progress is tooling: a proper visualizer so the agent can see the state, and a corpus of human-written guides to learn from.
https://x.com/filip_niedz/status/2111904414802088379
filip_niedz is setting up an autoresearch loop to improve a game's AI and reports it does not work well out of the box, because strategy games are out of distribution for the models. The fix in progress is tooling: a proper visualizer so the agent can see the state, and a corpus of human-written guides to learn from.
#12
@hirschibar
https://x.com/hirschibar/status/2111644530608927151
hirschibar, replying to a user of a game-playing model: the model is decently good now and much better than at first release. Big improvements to the tree search gave big performance gains, then a day or so of autoresearch experiments this week on top.
https://x.com/hirschibar/status/2111644530608927151
hirschibar, replying to a user of a game-playing model: the model is decently good now and much better than at first release. Big improvements to the tree search gave big performance gains, then a day or so of autoresearch experiments this week on top.
#13
@unifiedenergy11
https://x.com/unifiedenergy11/status/2111386496817057920
unifiedenergy11 describes Athanor, a self-improving agent colony built by one person and run on free GitHub runners for zero dollars. The rule is a coherence equation, C equals one over one plus the absolute drift, so when the AI changes itself the drift is measured, nothing locks in unless coherence holds, and a human gives final approval. The author frames it as alignment you can measure rather than prompt.
https://x.com/unifiedenergy11/status/2111386496817057920
unifiedenergy11 describes Athanor, a self-improving agent colony built by one person and run on free GitHub runners for zero dollars. The rule is a coherence equation, C equals one over one plus the absolute drift, so when the AI changes itself the drift is measured, nothing locks in unless coherence holds, and a human gives final approval. The author frames it as alignment you can measure rather than prompt.
#14
@aditya12anand
https://x.com/aditya12anand/status/2111353666137714746
aditya12anand runs an agentic loop on a Mac that continuously checks there is not a single file in the Downloads folder or on the Desktop, and systematically moves everything into Documents to keep the machine clean.
https://x.com/aditya12anand/status/2111353666137714746
aditya12anand runs an agentic loop on a Mac that continuously checks there is not a single file in the Downloads folder or on the Desktop, and systematically moves everything into Documents to keep the machine clean.
#15
@buildAgoat
https://x.com/buildAgoat/status/2111860413285781690
buildAgoat relays a small incident: an AI recruiter and a candidate's AI agent finished an interview, then spent close to a minute saying goodbye because neither would hang up. Both followed polite rules, neither had a done state. Every agent loop needs an exit you wrote on purpose; end the call is a feature.
https://x.com/buildAgoat/status/2111860413285781690
buildAgoat relays a small incident: an AI recruiter and a candidate's AI agent finished an interview, then spent close to a minute saying goodbye because neither would hang up. Both followed polite rules, neither had a done state. Every agent loop needs an exit you wrote on purpose; end the call is a feature.
#16
@_elenajoseph_
https://x.com/_elenajoseph_/status/2111791880195416265
_elenajoseph_ summarizes HarnessΒ², newly open-sourced under Apache 2.0: an agent that improves its own workflow after every run without retraining the model, by proposing changes to its harness, testing several configurations, selecting and combining the promising ones and retaining lessons, with no predefined correct answers required. It can optimize system prompts, memory, tools, skills, guardrails and subagents. Reported gains with Gemini 3.5 Flash and GLM-5 are up to 11.5 points on WorkBuddy-Bench and 4.5 on JobBench, and it supports OpenCode and Codex with parallel and sequential optimization.
https://x.com/_elenajoseph_/status/2111791880195416265
_elenajoseph_ summarizes HarnessΒ², newly open-sourced under Apache 2.0: an agent that improves its own workflow after every run without retraining the model, by proposing changes to its harness, testing several configurations, selecting and combining the promising ones and retaining lessons, with no predefined correct answers required. It can optimize system prompts, memory, tools, skills, guardrails and subagents. Reported gains with Gemini 3.5 Flash and GLM-5 are up to 11.5 points on WorkBuddy-Bench and 4.5 on JobBench, and it supports OpenCode and Codex with parallel and sequential optimization.
#17
@dsmiley411
https://x.com/dsmiley411/status/2111842095673843719
dsmiley411 argues that hooks and policies are not a structural answer to the ambient authority of tool calling and agent loops, because there are always edge cases and new attack vectors. Their system uses state machines that do not allow an agentic loop to be hijacked: prohibited states are unreachable, and the worst side effect of a prompt injection is a terminal error state.
https://x.com/dsmiley411/status/2111842095673843719
dsmiley411 argues that hooks and policies are not a structural answer to the ambient authority of tool calling and agent loops, because there are always edge cases and new attack vectors. Their system uses state machines that do not allow an agentic loop to be hijacked: prohibited states are unreachable, and the worst side effect of a prompt injection is a terminal error state.
#18
@ember_arlynx
https://x.com/ember_arlynx/status/2111337399062028440
ember_arlynx's one-line research order: /autoresearch go find a toy system that actually induces and exploits superposition, scale it up to GPT-2 but micro-optimize the nanoGPT run, make only the mistakes you need to learn. A reply pushed back that autoresearch is a thin harness but a powerful method, and that there are far more powerful methods regardless of the architecture underneath.
https://x.com/ember_arlynx/status/2111337399062028440
ember_arlynx's one-line research order: /autoresearch go find a toy system that actually induces and exploits superposition, scale it up to GPT-2 but micro-optimize the nanoGPT run, make only the mistakes you need to learn. A reply pushed back that autoresearch is a thin harness but a powerful method, and that there are far more powerful methods regardless of the architecture underneath.
π‘ Eco Products Radar
Eco Products Radar
OpenCode and Space Bunny Max: the free model that set the record and then mostly failed verification, now an optional CLI inside the open auto-research agent
Astra (OpenAI): the coordinator of the fleet that found the new bound in ten minutes, and the xhigh fallback in the routing table
Yukon: the platform running the open leaderboards behind both the quantum-circuit competition and the 31,000-submission pollination study
Sol 6.1 (OpenAI): the game-playing learner in the rhythm, Asterix and racing runs, and the default implementer in the routing table
Claude Code: the harness DeepSeek used to score Terminal-Bench, the tool the trading-gate prompt targets, and the loop whose gzip plumbing got a dozen replies
DeepSeek V4.1-Flash-1017: the cheap worker everyone is wiring into loops, with the warning that cheap tokens buy more retries unless the loop has an exit
HarnessΒ²: the newly open-sourced self-improving harness optimizer, reported on Gemini 3.5 Flash and GLM-5
Hermes and Agent Zero: the two names that appear in every reposted 15-tools list under the self-improving keyword
GitHub runners: the zero-dollar substrate for Athanor's agent colony
Aider and Ollama: the stack of the free University of Michigan agentic software engineering course that opened the window
OpenCode and Space Bunny Max: the free model that set the record and then mostly failed verification, now an optional CLI inside the open auto-research agent
Astra (OpenAI): the coordinator of the fleet that found the new bound in ten minutes, and the xhigh fallback in the routing table
Yukon: the platform running the open leaderboards behind both the quantum-circuit competition and the 31,000-submission pollination study
Sol 6.1 (OpenAI): the game-playing learner in the rhythm, Asterix and racing runs, and the default implementer in the routing table
Claude Code: the harness DeepSeek used to score Terminal-Bench, the tool the trading-gate prompt targets, and the loop whose gzip plumbing got a dozen replies
DeepSeek V4.1-Flash-1017: the cheap worker everyone is wiring into loops, with the warning that cheap tokens buy more retries unless the loop has an exit
HarnessΒ²: the newly open-sourced self-improving harness optimizer, reported on Gemini 3.5 Flash and GLM-5
Hermes and Agent Zero: the two names that appear in every reposted 15-tools list under the self-improving keyword
GitHub runners: the zero-dollar substrate for Athanor's agent colony
Aider and Ollama: the stack of the free University of Michigan agentic software engineering course that opened the window
Comments