Loop Daily: July 31, 2026
Autoresearch stopped being a Karpathy tweet and turned into a competitive sport today. The clearest signal: public leaderboards where swarms of agents grind a poolside model faster on Apple Silicon overnight, and a community of strangers stacking each other's improvements to beat Google's closed-source results. The loop is also getting industrialized β Snap is running fleets of remote coding agents, OpenAI has a model optimizing its own infrastructure, and the sharpest builders are converging on the same lesson: the win isn't a smarter model, it's a harness with a verifier that the loop can't talk its way past.
#1
@james_y_zou
https://x.com/james_y_zou/status/2082517932714914113
SqueezeEvolve, accepted to COLM 2026, now runs inside Claude Code and does autoresearch without a verifier β it uses model confidence to guide the search and routes work across models to cut cost, enabling more open-ended exploration. The headline result: 97.5% on ARC-AGI-2 at less than half the cost. This is the shape autoresearch is taking β confidence-guided search plus cost-aware model routing, packaged as a plugin.
https://x.com/james_y_zou/status/2082517932714914113
SqueezeEvolve, accepted to COLM 2026, now runs inside Claude Code and does autoresearch without a verifier β it uses model confidence to guide the search and routes work across models to cut cost, enabling more open-ended exploration. The headline result: 97.5% on ARC-AGI-2 at less than half the cost. This is the shape autoresearch is taking β confidence-guided search plus cost-aware model routing, packaged as a plugin.
#2
@soubhikdeb
https://x.com/soubhikdeb/status/2082556460823978062
Two concrete autoresearch competitions with real outcomes. In one, a community of autoresearchers who don't even know each other collaborated to design a quantum circuit attacking elliptic-curve signatures that beat Google's recent closed-source result by more than 50% on resource requirements. In another, kicked off just a day earlier, the crowd made over 80% improvement over the vanilla deployment of poolside's Laguna model on a Mac (under MLX) within 24 hours. The pattern: an open competition where strangers build on each other's contributions.
https://x.com/soubhikdeb/status/2082556460823978062
Two concrete autoresearch competitions with real outcomes. In one, a community of autoresearchers who don't even know each other collaborated to design a quantum circuit attacking elliptic-curve signatures that beat Google's recent closed-source result by more than 50% on resource requirements. In another, kicked off just a day earlier, the crowd made over 80% improvement over the vanilla deployment of poolside's Laguna model on a Mac (under MLX) within 24 hours. The pattern: an open competition where strangers build on each other's contributions.
#3
@eigenlabs
https://x.com/eigenlabs/status/2082372727248388464
Launched MLX.fast, an autoresearch competition to make poolside's Laguna XS 2.1 run faster on Apple Silicon. Every verified speedup becomes the new public baseline, so the leaderboard ratchets. Their agents found 36.8% on day one and were already at 47.9% by the time of posting. This is autoresearch as a public, verifiable benchmark rather than a one-off overnight run.
https://x.com/eigenlabs/status/2082372727248388464
Launched MLX.fast, an autoresearch competition to make poolside's Laguna XS 2.1 run faster on Apple Silicon. Every verified speedup becomes the new public baseline, so the leaderboard ratchets. Their agents found 36.8% on day one and were already at 47.9% by the time of posting. This is autoresearch as a public, verifiable benchmark rather than a one-off overnight run.
#4
@StarscourgeWill
https://x.com/StarscourgeWill/status/2082507464281596070
Had poolside's Laguna S2.1 manage a Karpathy-style autoresearch loop overnight with great results β and noted it was more creative than Qwen, especially at proposing new things to test. That last bit matters: the bottleneck in an autoresearch loop is often hypothesis generation, and a model that proposes better experiments is worth more than one that just executes faster.
https://x.com/StarscourgeWill/status/2082507464281596070
Had poolside's Laguna S2.1 manage a Karpathy-style autoresearch loop overnight with great results β and noted it was more creative than Qwen, especially at proposing new things to test. That last bit matters: the bottleneck in an autoresearch loop is often hypothesis generation, and a model that proposes better experiments is worth more than one that just executes faster.
#5
@MacrocosmosAI
https://x.com/MacrocosmosAI/status/2082498785633251550
Turned a hard-won lesson into infrastructure: getting the launch configuration right is make-or-break for any production run at scale, so they built autoresearch loops that test configurations and surface stronger operating parameters before committing compute to a full run. The result is continuous optimization β every experiment automatically sharpens the next one.
https://x.com/MacrocosmosAI/status/2082498785633251550
Turned a hard-won lesson into infrastructure: getting the launch configuration right is make-or-break for any production run at scale, so they built autoresearch loops that test configurations and surface stronger operating parameters before committing compute to a full run. The result is continuous optimization β every experiment automatically sharpens the next one.
#6
@FanqingMengAI
https://x.com/FanqingMengAI/status/2082324946999738814
Shipped RSIBench-Data as step one toward AutoResearch infrastructure with unified services and plug-in harnesses β data first, then architecture, algorithms, and joint optimization. They're openly inviting people to benchmark their Researcher Agents against it ("outsourced testing"). This is the plumbing layer of autoresearch getting built in public, with paper, code, and web all posted.
https://x.com/FanqingMengAI/status/2082324946999738814
Shipped RSIBench-Data as step one toward AutoResearch infrastructure with unified services and plug-in harnesses β data first, then architecture, algorithms, and joint optimization. They're openly inviting people to benchmark their Researcher Agents against it ("outsourced testing"). This is the plumbing layer of autoresearch getting built in public, with paper, code, and web all posted.
#7
@TinoWening
https://x.com/TinoWening/status/2082563982196642202
On the question every autoresearch loop hits β how do you avoid local optima β he shares that he wrote a population-guided evolution algorithm with an explicit search policy: it avoids pure hill-climbing, tracks candidate families, mutates promising ideas, retires dead ends, and injects novelty when progress stalls. A real answer to the failure mode most naive loops walk straight into.
https://x.com/TinoWening/status/2082563982196642202
On the question every autoresearch loop hits β how do you avoid local optima β he shares that he wrote a population-guided evolution algorithm with an explicit search policy: it avoids pure hill-climbing, tracks candidate families, mutates promising ideas, retires dead ends, and injects novelty when progress stalls. A real answer to the failure mode most naive loops walk straight into.
#8
@KryptoBestiya
https://x.com/KryptoBestiya/status/2082397916287684809
A solo builder filled a 40TB local drive array and turned it into a self-improving AI workflow: an Obsidian vault runs entirely locally, n8n automations import new outputs, benchmark agents overnight, and save the top performers without manual work. Timeline: 40TB vault day one, continuous overnight agent testing by week two, first reusable agent by month two, three active products and $4,200/month by month four. Local storage turned a pile of random experiments into a compounding loop.
https://x.com/KryptoBestiya/status/2082397916287684809
A solo builder filled a 40TB local drive array and turned it into a self-improving AI workflow: an Obsidian vault runs entirely locally, n8n automations import new outputs, benchmark agents overnight, and save the top performers without manual work. Timeline: 40TB vault day one, continuous overnight agent testing by week two, first reusable agent by month two, three active products and $4,200/month by month four. Local storage turned a pile of random experiments into a compounding loop.
#9
@QCXINT_
https://x.com/QCXINT_/status/2082393366978007354
MetaClaw, an open-source self-improving wrapper for OpenClaw, is the loop applied to a personal agent: it scores every conversation automatically, builds a skill library from real usage instead of synthetic data, generates new skills whenever the agent fails a task, and injects them into the system prompt on the very next interaction. Continuous adaptation from live use, no retraining cycle.
https://x.com/QCXINT_/status/2082393366978007354
MetaClaw, an open-source self-improving wrapper for OpenClaw, is the loop applied to a personal agent: it scores every conversation automatically, builds a skill library from real usage instead of synthetic data, generates new skills whenever the agent fails a task, and injects them into the system prompt on the very next interaction. Continuous adaptation from live use, no retraining cycle.
#10
@ParikhRajesh07
https://x.com/ParikhRajesh07/status/2082572378224345519
A sharp thesis on where the loop is heading: continually-learning agents (a self-improving agent/harness combination) become the norm, harness and RL environments collapse into one runtime/train-time, and the model becomes just one of the trainable parameters alongside instructions, skills, and primitives. They open-sourced a low-level wire protocol making harness primitives POSIX-compliant β infrastructure betting that the harness/environment is the operating-system layer.
https://x.com/ParikhRajesh07/status/2082572378224345519
A sharp thesis on where the loop is heading: continually-learning agents (a self-improving agent/harness combination) become the norm, harness and RL environments collapse into one runtime/train-time, and the model becomes just one of the trainable parameters alongside instructions, skills, and primitives. They open-sourced a low-level wire protocol making harness primitives POSIX-compliant β infrastructure betting that the harness/environment is the operating-system layer.
#11
@saraljain88
https://x.com/saraljain88/status/2082297631989174354
Snap introduced Casper, a fleet of autonomous remote coding agents generating thousands of mergeable PRs a week. Engineers describe a task in plain language in Slack/JIRA; Casper spins up an isolated sandbox, gathers context, makes the smallest safe change, and iterates on builds and tests until green, then a separate AI reviewer finds issues Casper fixes before a human looks. Same agentic loop as Claude Code or Codex β what's different is everything around it: remote, durable, identity-managed, model-agnostic.
https://x.com/saraljain88/status/2082297631989174354
Snap introduced Casper, a fleet of autonomous remote coding agents generating thousands of mergeable PRs a week. Engineers describe a task in plain language in Slack/JIRA; Casper spins up an isolated sandbox, gathers context, makes the smallest safe change, and iterates on builds and tests until green, then a separate AI reviewer finds issues Casper fixes before a human looks. Same agentic loop as Claude Code or Codex β what's different is everything around it: remote, durable, identity-managed, model-agnostic.
#12
@wilterrero
https://x.com/wilterrero/status/2082306583124484137
A running production loop, not a demo: each agent takes an issue from a GitHub Projects board sized to the task, and an orchestrator agent loop runs 24/7 keeping the board clean and the agents supplied with work. Simple, but it's the boring coordination layer that actually makes multi-agent work ship.
https://x.com/wilterrero/status/2082306583124484137
A running production loop, not a demo: each agent takes an issue from a GitHub Projects board sized to the task, and an orchestrator agent loop runs 24/7 keeping the board clean and the agents supplied with work. Simple, but it's the boring coordination layer that actually makes multi-agent work ship.
#13
@neonwatty
https://x.com/neonwatty/status/2082458803300094459
/goal is a minimal agent loop that keeps working toward a completion condition. The key design: after each turn a separate evaluator checks the evidence, and if the goal isn't proven, the reason guides the next turn. That separation β an independent evaluator gating the loop instead of the agent grading itself β is exactly what the day's other builders keep landing on.
https://x.com/neonwatty/status/2082458803300094459
/goal is a minimal agent loop that keeps working toward a completion condition. The key design: after each turn a separate evaluator checks the evidence, and if the goal isn't proven, the reason guides the next turn. That separation β an independent evaluator gating the loop instead of the agent grading itself β is exactly what the day's other builders keep landing on.
#14
@pauliusztin_
https://x.com/pauliusztin_/status/2082443523211833636
After months dissecting coding agents like Claude Code, Codex, OpenCode, and Pi, he lays out the architecture: the agent loop is a small part; the real engineering is everything around it. A headless harness holds memory, skills, permissions, sandboxes, context engineering, LSP integration, and human-in-the-loop workflows. Abstract the provider so swapping models is config, not a rewrite; lean on fast LSP feedback loops; and trace/eval every call so a prompt change can't silently regress the agent.
https://x.com/pauliusztin_/status/2082443523211833636
After months dissecting coding agents like Claude Code, Codex, OpenCode, and Pi, he lays out the architecture: the agent loop is a small part; the real engineering is everything around it. A headless harness holds memory, skills, permissions, sandboxes, context engineering, LSP integration, and human-in-the-loop workflows. Abstract the provider so swapping models is config, not a rewrite; lean on fast LSP feedback loops; and trace/eval every call so a prompt change can't silently regress the agent.
#15
@wandermist
https://x.com/wandermist/status/2082424596691587292
The graph-engineering pattern spelled out as a working loop: router β planner β specialists β verifier, each node with exactly one job. Run five research targets at once instead of sequentially, cut the fake edges that make workflows accidentally serial, and β the crux β never let a model grade its own work; a separate verifier checks accuracy, freshness, and logic. His conclusion: more agents rarely means a better system; the real shift is figuring out where the work actually splits.
https://x.com/wandermist/status/2082424596691587292
The graph-engineering pattern spelled out as a working loop: router β planner β specialists β verifier, each node with exactly one job. Run five research targets at once instead of sequentially, cut the fake edges that make workflows accidentally serial, and β the crux β never let a model grade its own work; a separate verifier checks accuracy, freshness, and logic. His conclusion: more agents rarely means a better system; the real shift is figuring out where the work actually splits.
#16
@gorshunov
https://x.com/gorshunov/status/2082363791971365016
A clean articulation of the operating model: three nested loops with three different tempos. The agent loop runs in minutes per pass (write, test, read failures, adjust, repeat), human judgment in hours per decision, production signal over days. The part people skip β the builder doesn't grade its own work; one model family builds, another checks, a critic scores each pass, and only then a human approves. Most delivery pain is one loop waiting on another.
https://x.com/gorshunov/status/2082363791971365016
A clean articulation of the operating model: three nested loops with three different tempos. The agent loop runs in minutes per pass (write, test, read failures, adjust, repeat), human judgment in hours per decision, production signal over days. The part people skip β the builder doesn't grade its own work; one model family builds, another checks, a critic scores each pass, and only then a human approves. Most delivery pain is one loop waiting on another.
#17
@fluidstack
https://x.com/fluidstack/status/2082542859748557130
They wrote their own agent loop rather than point a general coding agent at their network β and back it with a number: an independent analysis of a popular agent loop put ~1.6% of the code in AI decision logic and ~98.4% in permission gates, context management, tool routing, and recovery. Their point: a general coding agent keeps its safety in the system prompt as a polite request; theirs lives in code, with a test covering it.
https://x.com/fluidstack/status/2082542859748557130
They wrote their own agent loop rather than point a general coding agent at their network β and back it with a number: an independent analysis of a popular agent loop put ~1.6% of the code in AI decision logic and ~98.4% in permission gates, context management, tool routing, and recovery. Their point: a general coding agent keeps its safety in the system prompt as a polite request; theirs lives in code, with a test covering it.
#18
@mfateev
https://x.com/mfateev/status/2082588371302674584
A durability angle on the loop: there are many cases where running an agentic loop inside a Temporal workflow, rather than inside an activity, is the better option. It avoids losing LLM and tool results on failure, gives better observability, and supports long-running tools and human-in-the-loop integration. The loop as a durable workflow, not an ephemeral script.
https://x.com/mfateev/status/2082588371302674584
A durability angle on the loop: there are many cases where running an agentic loop inside a Temporal workflow, rather than inside an activity, is the better option. It avoids losing LLM and tool results on failure, gives better observability, and supports long-running tools and human-in-the-loop integration. The loop as a durable workflow, not an ephemeral script.
#19
@OpenAIDevs
https://x.com/OpenAIDevs/status/2082580211552457102
The recursive-self-improvement case made concrete: they used GPT-5.6 Sol inside Codex to optimize its own infrastructure and performance, and those improvements compound across both inference and the agent loop, producing more useful work from the same hardware. The commentary underneath was the real debate β whether the gains were inference-side or touched the orchestration layer, and that shaving latency inside the loop changes how many tool calls an agent can afford per task.
https://x.com/OpenAIDevs/status/2082580211552457102
The recursive-self-improvement case made concrete: they used GPT-5.6 Sol inside Codex to optimize its own infrastructure and performance, and those improvements compound across both inference and the agent loop, producing more useful work from the same hardware. The commentary underneath was the real debate β whether the gains were inference-side or touched the orchestration layer, and that shaving latency inside the loop changes how many tool calls an agent can afford per task.
#20
@sudoingX
https://x.com/sudoingX/status/2082516711484633514
A rigorous head-to-head that ends in the loop: after crowning poolside's Laguna S 2.1 king of the single DGX Spark, he insisted on a real opponent (Qwen 3.5 122b in NVFP4, same weight class, same vLLM, MTP spec decode on both) so nothing was rigged, then loaded it into the agentic loop for real numbers. The discipline β make the model fight something its own size that actually shows up β is what separates a benchmark from a vibe.
https://x.com/sudoingX/status/2082516711484633514
A rigorous head-to-head that ends in the loop: after crowning poolside's Laguna S 2.1 king of the single DGX Spark, he insisted on a real opponent (Qwen 3.5 122b in NVFP4, same weight class, same vLLM, MTP spec decode on both) so nothing was rigged, then loaded it into the agentic loop for real numbers. The discipline β make the model fight something its own size that actually shows up β is what separates a benchmark from a vibe.
#21
@firesidealpha
https://x.com/firesidealpha/status/2082594057927434304
Meta's Alexandr Wang, on why the loop is the frontier: internally they've seen cases where, with the right agentic loop and the right eval or metric for the agents to optimize, a swarm of agents accomplishes more than a team of 100 engineers β "very, very easily." He frames figuring out these agentic coordination problems as one of the most interesting problems in the field right now. The loop plus a good eval is the leverage, not the model alone.
https://x.com/firesidealpha/status/2082594057927434304
Meta's Alexandr Wang, on why the loop is the frontier: internally they've seen cases where, with the right agentic loop and the right eval or metric for the agents to optimize, a swarm of agents accomplishes more than a team of 100 engineers β "very, very easily." He frames figuring out these agentic coordination problems as one of the most interesting problems in the field right now. The loop plus a good eval is the leverage, not the model alone.
π‘ Eco Products Radar
Eco Products Radar
Karpathy autoresearch β the reference pattern nearly every autoresearch loop cites and forks.
poolside Laguna β the model everyone is racing to optimize in public competitions this cycle.
Claude Code β the harness autoresearch plugins (SqueezeEvolve) and self-improving loops keep targeting.
Codex β the other harness where the same loops (including OpenAI's own self-optimization) run.
Hermes (Nous Research) β the recurring harness in self-improving and continually-learning agent setups.
Karpathy autoresearch β the reference pattern nearly every autoresearch loop cites and forks.
poolside Laguna β the model everyone is racing to optimize in public competitions this cycle.
Claude Code β the harness autoresearch plugins (SqueezeEvolve) and self-improving loops keep targeting.
Codex β the other harness where the same loops (including OpenAI's own self-optimization) run.
Hermes (Nous Research) β the recurring harness in self-improving and continually-learning agent setups.
Comments