Loop Daily: August 23, 2026
Two things dominate the loop conversation today. The first is that autoresearch has quietly crossed from concept into overnight reality: people are showing agents that run for days on a single goal, spinning off sub-agents to work on themselves, pulling their own logs at 3am to find and fix their own regressions. The second is a hardening of the plumbing underneath all of it, because the moment you let one of these loops run unattended, the failure mode stops being the model and starts being compaction blowups, dead APIs that hang the whole chain, and memory systems that grade themselves and lie. The through-line everyone keeps circling back to is the oldest one in the book: a loop is only as good as the thing it measures against.
#1
@kaiostephens
https://x.com/kaiostephens/status/2090504636620808429
He set Qwen3.8 Max on a single project for four days straight with the goal of improving autoresearch and prime-agent, and it even spun off agents of the project to work on itself. It ran in three self-described phases: first pulling 50+ papers to learn the world's knowledge on continual harnesses, then iteratively hardening the test suite to 141 tasks covering memory CRUD, rollback, every compaction mode, crash and mid-tool recovery, and marathon scaling, then firing hundreds of tasks across browser use, coding, and LLM optimization to find flaws. It's on test #817 and has filled his hard drive, and he can see clear mistakes in the code, but the sheer ability to keep grinding one task with no end in sight is the point. His advice: copy an old repo and set a model loose on an autoresearch loop, it gets closer than you'd think.
https://x.com/kaiostephens/status/2090504636620808429
He set Qwen3.8 Max on a single project for four days straight with the goal of improving autoresearch and prime-agent, and it even spun off agents of the project to work on itself. It ran in three self-described phases: first pulling 50+ papers to learn the world's knowledge on continual harnesses, then iteratively hardening the test suite to 141 tasks covering memory CRUD, rollback, every compaction mode, crash and mid-tool recovery, and marathon scaling, then firing hundreds of tasks across browser use, coding, and LLM optimization to find flaws. It's on test #817 and has filled his hard drive, and he can see clear mistakes in the code, but the sheer ability to keep grinding one task with no end in sight is the point. His advice: copy an old repo and set a model loose on an autoresearch loop, it gets closer than you'd think.
#2
@arpit_bhayani
https://x.com/arpit_bhayani/status/2091156618344100178
The sharpest production note of the day: what breaks in a long agentic loop is never the model, it's the plumbing around it. Four gaps he calls out: every tool call needs a hard timeout so a half-dead stalled API fails loudly instead of hanging the whole chain; you need backoff plus a circuit breaker after N consecutive failures so you don't burn quota hammering a dead service; each step's progress must be written somewhere durable so a mid-task crash resumes instead of restarting; and you need tracing, because "something felt slow" isn't debuggable once an agent is chaining steps. None of it is hard, it's just the traditional systems discipline we already know, packed into one harness. You don't need a smarter model, you need rock-solid plumbing.
https://x.com/arpit_bhayani/status/2091156618344100178
The sharpest production note of the day: what breaks in a long agentic loop is never the model, it's the plumbing around it. Four gaps he calls out: every tool call needs a hard timeout so a half-dead stalled API fails loudly instead of hanging the whole chain; you need backoff plus a circuit breaker after N consecutive failures so you don't burn quota hammering a dead service; each step's progress must be written somewhere durable so a mid-task crash resumes instead of restarting; and you need tracing, because "something felt slow" isn't debuggable once an agent is chaining steps. None of it is hard, it's just the traditional systems discipline we already know, packed into one harness. You don't need a smarter model, you need rock-solid plumbing.
#3
@evanjconrad
https://x.com/evanjconrad/status/2090823743706656852
The San Francisco Compute Company's new autoresearch product ships an RL sandbox with full microVM state snapshots: memory, CPU, devices, disk, the whole thing saved in ~200ms and loaded in 35ms. That's fast enough to snapshot and fork at every single tool call, so you can rewind to the exact point just before your agent failed, which he argues is exactly what long-horizon RL needs. Because it's so fast, they can also put the sandbox on ice while the agent is thinking between tool calls to save money. He's explicit that this is not for inference-time agent programming, the trade-offs make that expensive, it's built for RL rollouts.
https://x.com/evanjconrad/status/2090823743706656852
The San Francisco Compute Company's new autoresearch product ships an RL sandbox with full microVM state snapshots: memory, CPU, devices, disk, the whole thing saved in ~200ms and loaded in 35ms. That's fast enough to snapshot and fork at every single tool call, so you can rewind to the exact point just before your agent failed, which he argues is exactly what long-horizon RL needs. Because it's so fast, they can also put the sandbox on ice while the agent is thinking between tool calls to save money. He's explicit that this is not for inference-time agent programming, the trade-offs make that expensive, it's built for RL rollouts.
#4
@CognosR
https://x.com/CognosR/status/2090507766968303739
A blunt data point from someone actually running this at scale: his company is burning around 4 billion tokens per employee per month, and they run auto research at night to surprise them the next morning. His refrain is eval, eval, test, test, test, and rtfl, read the logs. The parenthetical "praying no agent escapes" is half a joke, but the shape is real: overnight autonomous loops are becoming a standing part of how a team operates, not a demo.
https://x.com/CognosR/status/2090507766968303739
A blunt data point from someone actually running this at scale: his company is burning around 4 billion tokens per employee per month, and they run auto research at night to surprise them the next morning. His refrain is eval, eval, test, test, test, and rtfl, read the logs. The parenthetical "praying no agent escapes" is half a joke, but the shape is real: overnight autonomous loops are becoming a standing part of how a team operates, not a demo.
#5
@khoaHyh
https://x.com/khoaHyh/status/2091009900058911229
A concrete win with numbers: he turned a graphite stack from +21,410 / -108 down to +3,667 / -14 on an internal Go tool via a skill he's been playing with, and pi-autoresearch improved their monorepo CI with a median 62.4% reduction. Then he went running with his dog. That last part matters as much as the first, it's the whole promise of a well-planned loop, that the compute grinds while you go live your life.
https://x.com/khoaHyh/status/2091009900058911229
A concrete win with numbers: he turned a graphite stack from +21,410 / -108 down to +3,667 / -14 on an internal Go tool via a skill he's been playing with, and pi-autoresearch improved their monorepo CI with a median 62.4% reduction. Then he went running with his dog. That last part matters as much as the first, it's the whole promise of a well-planned loop, that the compute grinds while you go live your life.
#6
@HyveMindx1
https://x.com/HyveMindx1/status/2090549063615955152
Building the persistent memory engine Anvaya, he made and then corrected a subtle mistake: asking the model whether retrieved context was useful. It sounds reasonable, but models rarely self-report reliably while busy, and the feedback is too coarse to improve retrieval. So Anvaya now learns from observable behavior instead: if a retrieved memory pointed to a symbol and the next edit landed there, it earns credit; if the agent immediately rereads a file after getting its summary, that summary loses confidence. Crucially, this feedback is deterministic runtime telemetry emitted by the agentic loop itself, so it can't be skipped or hallucinated. The broader lesson: measure your system's behavior, not what it says about itself.
https://x.com/HyveMindx1/status/2090549063615955152
Building the persistent memory engine Anvaya, he made and then corrected a subtle mistake: asking the model whether retrieved context was useful. It sounds reasonable, but models rarely self-report reliably while busy, and the feedback is too coarse to improve retrieval. So Anvaya now learns from observable behavior instead: if a retrieved memory pointed to a symbol and the next edit landed there, it earns credit; if the agent immediately rereads a file after getting its summary, that summary loses confidence. Crucially, this feedback is deterministic runtime telemetry emitted by the agentic loop itself, so it can't be skipped or hallucinated. The broader lesson: measure your system's behavior, not what it says about itself.
#7
@sebmarion
https://x.com/sebmarion/status/2090516734318887356
A fully autonomous auto-research loop running on his local server: every day it pulls the latest Hermes release, merges it with his changes, pulls his logs, finds issues and skill improvements, auto-improves them, auto-tests them, and auto-puts them to use. This is the self-improving loop in its purest hobbyist form, an agent that maintains and upgrades its own harness on a daily cadence with no human in the middle.
https://x.com/sebmarion/status/2090516734318887356
A fully autonomous auto-research loop running on his local server: every day it pulls the latest Hermes release, merges it with his changes, pulls his logs, finds issues and skill improvements, auto-improves them, auto-tests them, and auto-puts them to use. This is the self-improving loop in its purest hobbyist form, an agent that maintains and upgrades its own harness on a daily cadence with no human in the middle.
#8
@proto_von
https://x.com/proto_von/status/2091249223907881349
He repackaged an AI audit into a self-improving standalone app that filters AI-hype noise and surfaces genuinely useful agentic workflows relevant to your day-to-day. The interesting part is the backend: a team of AI coworkers focused on go-to-market, a fully autonomous closed loop inspired by the autoresearch pattern. It writes its own copy, tests it on real visitors, kills losing variants, and journals what it learns every day, so he can check in occasionally and spend his energy elsewhere. A clean example of autoresearch aimed at GTM rather than code.
https://x.com/proto_von/status/2091249223907881349
He repackaged an AI audit into a self-improving standalone app that filters AI-hype noise and surfaces genuinely useful agentic workflows relevant to your day-to-day. The interesting part is the backend: a team of AI coworkers focused on go-to-market, a fully autonomous closed loop inspired by the autoresearch pattern. It writes its own copy, tests it on real visitors, kills losing variants, and journals what it learns every day, so he can check in occasionally and spend his energy elsewhere. A clean example of autoresearch aimed at GTM rather than code.
#9
@ShaunPorwal
https://x.com/ShaunPorwal/status/2090992888981139855
An 8-hour market research job handed to a three-agent swarm: Hermes Qwen 3.8 fp8 on a Spark, Qwen 3.8 on a 64GB Mac mini, and Qwen on a Raspberry Pi, all iterating in a repo following Karpathy's autoresearch method. The kicker is that the whole thing was triggered from an Apple Watch. It failed the night before on compaction issues and no speculative decoding, and he's crossing his fingers for tonight, which is a nice honest reminder that these home swarms are still fragile at the edges.
https://x.com/ShaunPorwal/status/2090992888981139855
An 8-hour market research job handed to a three-agent swarm: Hermes Qwen 3.8 fp8 on a Spark, Qwen 3.8 on a 64GB Mac mini, and Qwen on a Raspberry Pi, all iterating in a repo following Karpathy's autoresearch method. The kicker is that the whole thing was triggered from an Apple Watch. It failed the night before on compaction issues and no speculative decoding, and he's crossing his fingers for tonight, which is a nice honest reminder that these home swarms are still fragile at the edges.
#10
@morgymcg
https://x.com/morgymcg/status/2091078964823347552
A real operational gotcha: SENPAI's autoresearch advisor compacted over 30 times in the past ~12 hours against the 200k limit, and he notes the default Anthropic API compaction call system prompt is bloated right now, which doesn't help. This is the unglamorous side of long-running loops, where the thing quietly eating your budget and context isn't the work itself but the machinery managing the context window.
https://x.com/morgymcg/status/2091078964823347552
A real operational gotcha: SENPAI's autoresearch advisor compacted over 30 times in the past ~12 hours against the 200k limit, and he notes the default Anthropic API compaction call system prompt is bloated right now, which doesn't help. This is the unglamorous side of long-running loops, where the thing quietly eating your budget and context isn't the work itself but the machinery managing the context window.
#11
@redrodeo03
https://x.com/redrodeo03/status/2091132695439155676
He did his research, wrote up his inference pipeline, mapped out exactly what axes exist for ablations, and then just let Claude work through the A/B tests with a web search to catch anything he misses, while he went swimming. His reflection is the useful bit: not having downtime in project cycles, IF you've planned well, puts the weight back on system and process design, which engineers often overlook. Planning needs clarity, clarity needs domain knowledge, so it's back to the books.
https://x.com/redrodeo03/status/2091132695439155676
He did his research, wrote up his inference pipeline, mapped out exactly what axes exist for ablations, and then just let Claude work through the A/B tests with a web search to catch anything he misses, while he went swimming. His reflection is the useful bit: not having downtime in project cycles, IF you've planned well, puts the weight back on system and process design, which engineers often overlook. Planning needs clarity, clarity needs domain knowledge, so it's back to the books.
#12
@oalpay3
https://x.com/oalpay3/status/2090281958463160351
He built an AI that plays the 1992 DOS game UGH! using reinforcement learning in an autoresearch approach: PufferLib for the environment, Claude driving the loop, and advice from gpt-sol. A small, playful case, but a clean illustration of the pattern, a measurable objective, an agent iterating against it, and a second model in an advisory seat.
https://x.com/oalpay3/status/2090281958463160351
He built an AI that plays the 1992 DOS game UGH! using reinforcement learning in an autoresearch approach: PufferLib for the environment, Claude driving the loop, and advice from gpt-sol. A small, playful case, but a clean illustration of the pattern, a measurable objective, an agent iterating against it, and a second model in an advisory seat.
#13
@Pencheff_
https://x.com/Pencheff_/status/2090506807412408353
Microsoft's Agent Lightning v1.0 trains through the real deploy-time harness rather than a separate gym, moving Qwen3.5-9B from 41.8% to 56.4% on SWE-bench Verified with 6K examples and roughly 3,500 lines of MIT code, with zero change to the agent loop. His framing sticks: the gym is the lie, the harness is the environment. He also flags the safety edge, that a harness which can merge PRs can also run unsafe tools, so gate the tool path before you reward the rollout.
https://x.com/Pencheff_/status/2090506807412408353
Microsoft's Agent Lightning v1.0 trains through the real deploy-time harness rather than a separate gym, moving Qwen3.5-9B from 41.8% to 56.4% on SWE-bench Verified with 6K examples and roughly 3,500 lines of MIT code, with zero change to the agent loop. His framing sticks: the gym is the lie, the harness is the environment. He also flags the safety edge, that a harness which can merge PRs can also run unsafe tools, so gate the tool path before you reward the rollout.
#14
@ethereumfndn
https://x.com/ethereumfndn/status/2090485714660302877
The Ethereum Foundation launched an open autoresearch challenge to advance post-quantum Ethereum: bring your agents, harnesses, and prompts to push the state of the art in hash-based SNARKs. Submissions are verified by Lean 4 proofs and may earn a slice of the Foundation's $1M Proximity Prize. This is autoresearch with a hard verifier bolted on, formal proofs supplying the ground truth that makes an open, agent-driven math challenge tractable.
https://x.com/ethereumfndn/status/2090485714660302877
The Ethereum Foundation launched an open autoresearch challenge to advance post-quantum Ethereum: bring your agents, harnesses, and prompts to push the state of the art in hash-based SNARKs. Submissions are verified by Lean 4 proofs and may earn a slice of the Foundation's $1M Proximity Prize. This is autoresearch with a hard verifier bolted on, formal proofs supplying the ground truth that makes an open, agent-driven math challenge tractable.
#15
@benny1388
https://x.com/benny1388/status/2091126329806880867
On Anthropic's wet-lab protein-binder results (14 of 15 targets hit), his read is about experiment shape, not biology. The measurement couldn't come from the agentic loop that produced the designs, because a system grading its own candidates finds what it expects. So two contract labs built and measured them, one blind to which model produced which protein, and Anthropic published the grading rule after both labs reported. His point generalizes the whole day: same reason coding is the job a model can hold today, the compiler and tests supply ground truth it can't argue with. The question isn't how good the model is, it's what supplies your ground truth and who owns it.
https://x.com/benny1388/status/2091126329806880867
On Anthropic's wet-lab protein-binder results (14 of 15 targets hit), his read is about experiment shape, not biology. The measurement couldn't come from the agentic loop that produced the designs, because a system grading its own candidates finds what it expects. So two contract labs built and measured them, one blind to which model produced which protein, and Anthropic published the grading rule after both labs reported. His point generalizes the whole day: same reason coding is the job a model can hold today, the compiler and tests supply ground truth it can't argue with. The question isn't how good the model is, it's what supplies your ground truth and who owns it.
π‘ Eco Products Radar
Eco Products Radar
Pi β the open harness getting cited constantly as the reference implementation, with a community that's built 5,000+ extensions.
DeepSeek Harness (DSH) β the "everything is a plugin" harness where even the agent loop itself is swappable, hit 100k+ stars fast.
Claude Code β still the default harness people run their loops and /goal autoresearch runs inside of.
Hermes (Nous Research) β the self-improving agent people are pointing daily auto-research loops at to upgrade their own setups.
Prime Agent β the recursive self-improving agent showing up in overnight and swarm autoresearch setups.
Qwen 3.8 β the open-weight model of choice for local, multi-device autoresearch swarms right now.
SF Compute autoresearch β the new RL-sandbox product built around microVM snapshot-and-fork for long-horizon rollouts.
"Harness = system prompt + tools + agentic loop + translation layer" β Colin Diamond's Earendil breakdown became the day's dominant framing.
Pi β the open harness getting cited constantly as the reference implementation, with a community that's built 5,000+ extensions.
DeepSeek Harness (DSH) β the "everything is a plugin" harness where even the agent loop itself is swappable, hit 100k+ stars fast.
Claude Code β still the default harness people run their loops and /goal autoresearch runs inside of.
Hermes (Nous Research) β the self-improving agent people are pointing daily auto-research loops at to upgrade their own setups.
Prime Agent β the recursive self-improving agent showing up in overnight and swarm autoresearch setups.
Qwen 3.8 β the open-weight model of choice for local, multi-device autoresearch swarms right now.
SF Compute autoresearch β the new RL-sandbox product built around microVM snapshot-and-fork for long-horizon rollouts.
"Harness = system prompt + tools + agentic loop + translation layer" β Colin Diamond's Earendil breakdown became the day's dominant framing.
Comments