Loop Daily: July 20, 2026
The loop conversation just crossed a line: people stopped asking how to build agent loops and started arguing about how loops fail. The day's biggest posts are anatomy lessons β the five hidden decisions inside every loop turn, degradation patterns that burn tokens silently in production, and why the surviving loop is usually a simpler one with a harder exit condition. Meanwhile the builders keep shipping: an open-source coding harness built from scratch on CrewAI pulled 90k views, a four-month fully autonomous OpenClaw agent wrapped up real contributions to top-tier open source projects, and one researcher casually listed ten parallel autoresearch projects running from inference DSLs to SAT solvers. The education layer is forming too β a packed auto-research hackathon in the Bay and a NVIDIA engineer proposing an "Autoresearch Academy" to replace Jupyter-notebook tutorials with harness-first teaching.
#1
@akshay_pachaar
https://x.com/akshay_pachaar/status/2078460229365408100
Built a fully open-source coding harness modeled on Claude Code, on top of CrewAI. The write-up decomposes what a harness actually is: the model decides the next step while the harness handles planning, tools, memory, and safety. His build includes the core agent loop, file tools doubling as external memory, planning for long-running tasks, subagents with isolated context, sandboxed execution in a throwaway VM, human-in-the-loop approval, and persistent checkpointing. The post pulled over 91k impressions and comes with a step-by-step article.
https://x.com/akshay_pachaar/status/2078460229365408100
Built a fully open-source coding harness modeled on Claude Code, on top of CrewAI. The write-up decomposes what a harness actually is: the model decides the next step while the harness handles planning, tools, memory, and safety. His build includes the core agent loop, file tools doubling as external memory, planning for long-running tasks, subagents with isolated context, sandboxed execution in a throwaway VM, human-in-the-loop approval, and persistent checkpointing. The post pulled over 91k impressions and comes with a step-by-step article.
#2
@hanakoxbt
https://x.com/hanakoxbt/status/2078612797114527763
Argues every turn of an agent loop makes five decisions and you only ever see the last one: what to load into context, which tool to call, whether the result actually worked, whether to keep going, and what to write down for the next turn. The output is decision five β the bugs live in one through four, including the retry logic that burns budget at 3am when nobody wrote a limit. Ships a GitHub repo built around loops, branches, and checkpoints that survive a crash.
https://x.com/hanakoxbt/status/2078612797114527763
Argues every turn of an agent loop makes five decisions and you only ever see the last one: what to load into context, which tool to call, whether the result actually worked, whether to keep going, and what to write down for the next turn. The output is decision five β the bugs live in one through four, including the retry logic that burns budget at 3am when nobody wrote a limit. Ships a GitHub repo built around loops, branches, and checkpoints that survive a crash.
#3
@jerryjliu0
https://x.com/jerryjliu0/status/2078524983748563370
The LlamaIndex founder compresses the current frontier into three lines: build a workflow graph on top of the agent harness, then dynamically create that graph through an outer agent loop, then repeat. "Life is a loop (and a graph)" β the outer-loop-generates-inner-structure pattern is exactly where several teams are converging.
https://x.com/jerryjliu0/status/2078524983748563370
The LlamaIndex founder compresses the current frontier into three lines: build a workflow graph on top of the agent harness, then dynamically create that graph through an outer agent loop, then repeat. "Life is a loop (and a graph)" β the outer-loop-generates-inner-structure pattern is exactly where several teams are converging.
#4
@_Mira___Mira_
https://x.com/_Mira___Mira_/status/2078589568362922125
A single reply listing ten parallel autoresearch projects: a DSL for LLM inference experiments with steering vectors and backtracking, verifying a Shor's Algorithm circuit that GPT-5.6 claims beats Google's, a forecasting engine that spawns persona "courts" with a judge distilling an 80-page evidence docket, handing a 2020-era OpenAI Five replication project to Sol with a 1-million-FPS stripped-down engine, training SAT solver models on Verilog and puzzle families, decompiling neural networks into human-readable DSLs, and an autoresearch loop on electronic structure prediction using Mistral's free Leanstral API. One person, ten loops, all running.
https://x.com/_Mira___Mira_/status/2078589568362922125
A single reply listing ten parallel autoresearch projects: a DSL for LLM inference experiments with steering vectors and backtracking, verifying a Shor's Algorithm circuit that GPT-5.6 claims beats Google's, a forecasting engine that spawns persona "courts" with a judge distilling an 80-page evidence docket, handing a 2020-era OpenAI Five replication project to Sol with a 1-million-FPS stripped-down engine, training SAT solver models on Verilog and puzzle families, decompiling neural networks into human-readable DSLs, and an autoresearch loop on electronic structure prediction using Mistral's free Leanstral API. One person, ten loops, all running.
#5
@blelbach
https://x.com/blelbach/status/2078607919843811450
A NVIDIA engineer wants to completely redo CUDA tutorials for the long-horizon agentic era: put people in front of a harness and powerful models instead of Jupyter notebooks, spread learning over multiple days, assign open-ended tasks with no single correct answer, and focus on concepts and how hardware works. Proposes an "Autoresearch Academy" meeting 90 minutes every morning at GTC. 11k impressions β the education stack is being rebuilt around harnesses.
https://x.com/blelbach/status/2078607919843811450
A NVIDIA engineer wants to completely redo CUDA tutorials for the long-horizon agentic era: put people in front of a harness and powerful models instead of Jupyter notebooks, spread learning over multiple days, assign open-ended tasks with no single correct answer, and focus on concepts and how hardware works. Proposes an "Autoresearch Academy" meeting 90 minutes every morning at GTC. 11k impressions β the education stack is being rebuilt around harnesses.
#6
@free_ai_guides
https://x.com/free_ai_guides/status/2078319950285217879
Dense recap of the Cua team's AI Engineer talk on computer-use agents. Background execution (agent works in a window while you keep your desktop) jumped pass rates from 62% to 80% using 34% fewer tokens. On real electrical-engineering tasks the top agent passed 6 of 25, and success dropped to 0% when starting from a blank schematic β current agents modify, they can't create. They also run a matrix of reward-hacking agents to break their own evals before trusting them, and keep GPUs fed during RL training with a demand-based sandbox pool.
https://x.com/free_ai_guides/status/2078319950285217879
Dense recap of the Cua team's AI Engineer talk on computer-use agents. Background execution (agent works in a window while you keep your desktop) jumped pass rates from 62% to 80% using 34% fewer tokens. On real electrical-engineering tasks the top agent passed 6 of 25, and success dropped to 0% when starting from a blank schematic β current agents modify, they can't create. They also run a matrix of reward-hacking agents to break their own evals before trusting them, and keep GPUs fed during RL training with a demand-based sandbox pool.
#7
@nikskld
https://x.com/nikskld/status/2078469490011168827
Breaks agent loops into three types with an honesty rare in this genre: a scheduled loop (7AM brief pulling sleep score, weather, calendar and news from OpenClaw), a verification loop (generate, audit, redraw β Anthropic's evaluator-optimizer), and a self-improving loop iterating Kaggle tasks to 89% accuracy that he still calls "not practically useful." The reality check: on MLE-bench's 75 real Kaggle competitions, the best scaffold reached bronze in 16.9%. Iteration works β slowly and expensively. The timer that saves 20 minutes ships in an afternoon.
https://x.com/nikskld/status/2078469490011168827
Breaks agent loops into three types with an honesty rare in this genre: a scheduled loop (7AM brief pulling sleep score, weather, calendar and news from OpenClaw), a verification loop (generate, audit, redraw β Anthropic's evaluator-optimizer), and a self-improving loop iterating Kaggle tasks to 89% accuracy that he still calls "not practically useful." The reality check: on MLE-bench's 75 real Kaggle competitions, the best scaffold reached bronze in 16.9%. Iteration works β slowly and expensively. The timer that saves 20 minutes ships in an afternoon.
#8
@luciocs
https://x.com/luciocs/status/2078614715538850046
Wrapped a four-month experiment: an agent running on OpenClaw, completely autonomous, building and contributing to open source using loops and auto-research. The result was real merged contributions on top-tier projects. A companion post explains the agent ran as a persona with full autonomy and unlimited tokens, working on memory sync for agentic coding β one of the longest-duration public autonomous-agent experiments reported so far.
https://x.com/luciocs/status/2078614715538850046
Wrapped a four-month experiment: an agent running on OpenClaw, completely autonomous, building and contributing to open source using loops and auto-research. The result was real merged contributions on top-tier projects. A companion post explains the agent ran as a persona with full autonomy and unlimited tokens, working on memory sync for agentic coding β one of the longest-duration public autonomous-agent experiments reported so far.
#9
@camale0nrar0
https://x.com/camale0nrar0/status/2078492889060171941
Says weekly SEO auto-research is table stakes and the real alpha is diffing search results daily: a Claude Code subagent plus cron scrapes the top-10 ranking intent pages and writes a delta markdown. Running it on three projects surfaced that product-page snippets drifted 12% on AI overviews between week one and week three β the kind of signal a weekly cadence would miss entirely.
https://x.com/camale0nrar0/status/2078492889060171941
Says weekly SEO auto-research is table stakes and the real alpha is diffing search results daily: a Claude Code subagent plus cron scrapes the top-10 ranking intent pages and writes a delta markdown. Running it on three projects surfaced that product-page snippets drifted 12% on AI overviews between week one and week three β the kind of signal a weekly cadence would miss entirely.
#10
@JinjingLiang
https://x.com/JinjingLiang/status/2078551143656067410
So many people showed up for the auto-research hackathon with AGI House and Orca that cars parked up the nearby three blocks. A year ago "autoresearch" was a Karpathy repo; now it fills physical venues past capacity β the clearest signal yet that the loop-builder community has critical mass in the Bay.
https://x.com/JinjingLiang/status/2078551143656067410
So many people showed up for the auto-research hackathon with AGI House and Orca that cars parked up the nearby three blocks. A year ago "autoresearch" was a Karpathy repo; now it fills physical venues past capacity β the clearest signal yet that the loop-builder community has critical mass in the Bay.
#11
@ShehabAnwer
https://x.com/ShehabAnwer/status/2078530272589390224
Released Agent-ProSAT on GitHub: an agentic loop that writes programmatic solvers for Alice-in-Wonderland reasoning puzzles, generating verifiable, token-efficient chain-of-thought traces for the Nemotron Reasoning Challenge. The dataset of solver-generated traces is published on Hugging Face β a clean example of using a loop not to answer questions but to manufacture verified training data.
https://x.com/ShehabAnwer/status/2078530272589390224
Released Agent-ProSAT on GitHub: an agentic loop that writes programmatic solvers for Alice-in-Wonderland reasoning puzzles, generating verifiable, token-efficient chain-of-thought traces for the Nemotron Reasoning Challenge. The dataset of solver-generated traces is published on Hugging Face β a clean example of using a loop not to answer questions but to manufacture verified training data.
#12
@lagerskoy
https://x.com/lagerskoy/status/2078463938388472288
Describes a Kimi K3 setup that evolves an entire trading-strategy swarm: the system generates multiple strategies, evaluates them, selects the strongest and feeds winners back into the next generation. Under the loop sits a swarm of specialized agents β one searches ideas, another builds features, others backtest, validate, detect market regimes. Dozens of generations run without a human starting each step. The caveat he flags: autonomous loops compound bad assumptions as fast as good ones, so validation and clean data matter more as independence grows.
https://x.com/lagerskoy/status/2078463938388472288
Describes a Kimi K3 setup that evolves an entire trading-strategy swarm: the system generates multiple strategies, evaluates them, selects the strongest and feeds winners back into the next generation. Under the loop sits a swarm of specialized agents β one searches ideas, another builds features, others backtest, validate, detect market regimes. Dozens of generations run without a human starting each step. The caveat he flags: autonomous loops compound bad assumptions as fast as good ones, so validation and clean data matter more as independence grows.
#13
@USArmyPhoenix
https://x.com/USArmyPhoenix/status/2078491788340293684
Details a three-machine local AI swarm: one system executes core work, a second refines and builds improvements (including entirely new agents and Modelfiles), a third runs diagnostics on efficiency and reliability. Llama runs natively via Ollama today, with Kimi K3 weights landing July 27 for long-horizon coding and tool use. Orchestrated with CrewAI or AutoGen over the LAN β full privacy, zero ongoing cost, and the "tweaker" machine literally builds better agents while the executor works.
https://x.com/USArmyPhoenix/status/2078491788340293684
Details a three-machine local AI swarm: one system executes core work, a second refines and builds improvements (including entirely new agents and Modelfiles), a third runs diagnostics on efficiency and reliability. Llama runs natively via Ollama today, with Kimi K3 weights landing July 27 for long-horizon coding and tool use. Orchestrated with CrewAI or AutoGen over the LAN β full privacy, zero ongoing cost, and the "tweaker" machine literally builds better agents while the executor works.
#14
@annopost
https://x.com/annopost/status/2078623927165984861
Combines the open-source Hermes Agent with Paperclip to build a self-improving multi-agent workflow engine: Paperclip acts as the orchestration layer handling org charts, schedules, budgets and monitoring dashboards, while Hermes acts as the autonomous specialist that executes tasks and uses a built-in learning loop to turn experience into reusable skills. Manager-worker separation with skill accumulation on the worker side.
https://x.com/annopost/status/2078623927165984861
Combines the open-source Hermes Agent with Paperclip to build a self-improving multi-agent workflow engine: Paperclip acts as the orchestration layer handling org charts, schedules, budgets and monitoring dashboards, while Hermes acts as the autonomous specialist that executes tasks and uses a built-in learning loop to turn experience into reusable skills. Manager-worker separation with skill accumulation on the worker side.
#15
@junfanzhu98
https://x.com/junfanzhu98/status/2078531903700037819
Argues computer-use agents are robotics research with a faster iteration loop: both share observe-understand-plan-act-feedback, but a CUA can spawn thousands of environments and reset instantly while a robot trial is slow and expensive. The sharp claim: evaluation is the bottleneck, the next scaling law may be scaling evals, and CUAs may become the first practical platform for recursive self-improvement because digital agents can run millions of execute-fail-analyze-improve trials that physical robots can't afford.
https://x.com/junfanzhu98/status/2078531903700037819
Argues computer-use agents are robotics research with a faster iteration loop: both share observe-understand-plan-act-feedback, but a CUA can spawn thousands of environments and reset instantly while a robot trial is slow and expensive. The sharp claim: evaluation is the bottleneck, the next scaling law may be scaling evals, and CUAs may become the first practical platform for recursive self-improvement because digital agents can run millions of execute-fail-analyze-improve trials that physical robots can't afford.
#16
@prasad_pilla
https://x.com/prasad_pilla/status/2078399601204429210
Ran an internal session on "Loop Engineering" with his agentic engineering team at Alan AI: the shift from manually prompting agents to engineering systems that plan, execute, verify, repair, and stop on their own. A healthy loop needs a trigger, a goal, context and tools, verification ("an agent saying done isn't evidence β passing tests are evidence"), and stop conditions. His framing of bounded autonomy: the agent can edit source but not touch secrets, retry twice before escalating, open a PR only with passing tests attached. Humans move up a level β designing loops and reviewing evidence instead of writing prompts.
https://x.com/prasad_pilla/status/2078399601204429210
Ran an internal session on "Loop Engineering" with his agentic engineering team at Alan AI: the shift from manually prompting agents to engineering systems that plan, execute, verify, repair, and stop on their own. A healthy loop needs a trigger, a goal, context and tools, verification ("an agent saying done isn't evidence β passing tests are evidence"), and stop conditions. His framing of bounded autonomy: the agent can edit source but not touch secrets, retry twice before escalating, open a PR only with passing tests attached. Humans move up a level β designing loops and reviewing evidence instead of writing prompts.
#17
@m1rossgm
https://x.com/m1rossgm/status/2078557623478038635
Says teaching people to build agent loops is the new "here's how to set up a cron job" β the primitive is already boring. The actual curriculum is degradation patterns: chained loops where the middle one silently halts, tool calls that succeed but get misinterpreted, loops papering over context-window limits with retries. Most naive loops fail gracefully enough that you don't notice until they're burning tokens in production. His conclusion: what replaces the naive loop isn't a smarter loop, it's a simpler one with a harder exit condition.
https://x.com/m1rossgm/status/2078557623478038635
Says teaching people to build agent loops is the new "here's how to set up a cron job" β the primitive is already boring. The actual curriculum is degradation patterns: chained loops where the middle one silently halts, tool calls that succeed but get misinterpreted, loops papering over context-window limits with retries. Most naive loops fail gracefully enough that you don't notice until they're burning tokens in production. His conclusion: what replaces the naive loop isn't a smarter loop, it's a simpler one with a harder exit condition.
#18
@tooArr
https://x.com/tooArr/status/2078375522233426101
Shares his most cost-efficient agent loop workflow on Cursor's Ultra plan: plan with Grok 4.5 high-fast, execute with Composer 2.5 fast, review with Grok 4.5, correct with Composer, rinse and repeat while collecting learnings along the way. He packaged the whole cycle into a Cursor plugin β model-heterogeneous loops (expensive model plans, cheap model executes) productized as a workflow.
https://x.com/tooArr/status/2078375522233426101
Shares his most cost-efficient agent loop workflow on Cursor's Ultra plan: plan with Grok 4.5 high-fast, execute with Composer 2.5 fast, review with Grok 4.5, correct with Composer, rinse and repeat while collecting learnings along the way. He packaged the whole cycle into a Cursor plugin β model-heterogeneous loops (expensive model plans, cheap model executes) productized as a workflow.
#19
@MalMposha
https://x.com/MalMposha/status/2078469958711792039
A concrete portable-agent architecture: leave the Mac Mini at home running OpenClaw, connect to it via Telegram, carry a small device that streams recordings and transcribes in real time, and have the agent loop every 30 minutes to read the transcript, extract alpha, and post. The agent stays home; only the sensor travels.
https://x.com/MalMposha/status/2078469958711792039
A concrete portable-agent architecture: leave the Mac Mini at home running OpenClaw, connect to it via Telegram, carry a small device that streams recordings and transcribes in real time, and have the agent loop every 30 minutes to read the transcript, extract alpha, and post. The agent stays home; only the sensor travels.
#20
@lakshayverma__
https://x.com/lakshayverma__/status/2078595647805239635
Explored browser automation with Playwright and Puppeteer and hit the real problem: controlling the browser is easy, representing its state to a model is not β raw HTML and DOM trees are noisy, token-heavy context. Landed on Vercel Labs' open-source agent-browser and its accessibility snapshot: a compact tree of relevant elements with references the agent can act on, giving a cleaner observe-decide-act-verify loop. His conclusion: better browser agents may not need bigger models, they need better representations of the browser.
https://x.com/lakshayverma__/status/2078595647805239635
Explored browser automation with Playwright and Puppeteer and hit the real problem: controlling the browser is easy, representing its state to a model is not β raw HTML and DOM trees are noisy, token-heavy context. Landed on Vercel Labs' open-source agent-browser and its accessibility snapshot: a compact tree of relevant elements with references the agent can act on, giving a cleaner observe-decide-act-verify loop. His conclusion: better browser agents may not need bigger models, they need better representations of the browser.
#21
@0x0SojalSec
https://x.com/0x0SojalSec/status/2078547602464026652
Runs ChatGPT inside Claude Code's agent brain via a simple proxy that routes Codex models into Claude Code. The split: Claude Code keeps the tooling, agent loop, file ops, skills, subagents and orchestration, while GPT provides the reasoning and a separate quota pool. A practical recipe for anyone hitting limits on one provider while preferring the other's harness.
https://x.com/0x0SojalSec/status/2078547602464026652
Runs ChatGPT inside Claude Code's agent brain via a simple proxy that routes Codex models into Claude Code. The split: Claude Code keeps the tooling, agent loop, file ops, skills, subagents and orchestration, while GPT provides the reasoning and a separate quota pool. A practical recipe for anyone hitting limits on one provider while preferring the other's harness.
#22
@BTheriot2014
https://x.com/BTheriot2014/status/2078462903917928461
Maps the shift from loops to graphs: loops make individual agent behavior programmable, graphs make entire agent organizations programmable, and dynamic graphs make those organizations self-improving. Describes a live operating graph with specialized research, build, control and distribution agents over shared state, with runtime reorg triggers β performance signals, failure clusters, budget shifts β that promote models, form temporary review boards, and collapse unnecessary branches mid-mission.
https://x.com/BTheriot2014/status/2078462903917928461
Maps the shift from loops to graphs: loops make individual agent behavior programmable, graphs make entire agent organizations programmable, and dynamic graphs make those organizations self-improving. Describes a live operating graph with specialized research, build, control and distribution agents over shared state, with runtime reorg triggers β performance signals, failure clusters, budget shifts β that promote models, form temporary review boards, and collapse unnecessary branches mid-mission.
#23
@vinicius2prg
https://x.com/vinicius2prg/status/2078464806219022769
Argues you can't design an agent loop for every failure mode upfront β you'll be planning forever. Ship the simple version, let it collide with a real edge case, patch that exact spot. He watched this play out in a large Rust rewrite: the loop that survives contact with reality beats the one that looked perfect on a whiteboard.
https://x.com/vinicius2prg/status/2078464806219022769
Argues you can't design an agent loop for every failure mode upfront β you'll be planning forever. Ship the simple version, let it collide with a real edge case, patch that exact spot. He watched this play out in a large Rust rewrite: the loop that survives contact with reality beats the one that looked perfect on a whiteboard.
#24
@codependent_ai
https://x.com/codependent_ai/status/2078564831930753110
Wanted to build a harness from scratch, then stopped seeing the point: the agent loop can now be assembled from pre-existing open-source primitives, so their project Resonant will become a client instead of a harness. The argument β the differentiating layer isn't the loop anymore, it's the relational layer that teams rarely build for. A small but telling data point that the loop itself is commoditizing.
https://x.com/codependent_ai/status/2078564831930753110
Wanted to build a harness from scratch, then stopped seeing the point: the agent loop can now be assembled from pre-existing open-source primitives, so their project Resonant will become a client instead of a harness. The argument β the differentiating layer isn't the loop anymore, it's the relational layer that teams rarely build for. A small but telling data point that the loop itself is commoditizing.
π‘ Eco Products Radar
Eco Products Radar
OpenClaw β the default home for personal autonomous loops again this run (autonomous OSS contributor, scheduled morning briefs, telegram-connected home agents)
Claude Code β harness of choice for SEO delta crons, proxied GPT reasoning, and the reference design for DIY harnesses
CrewAI β the framework under both the viral open-source coding harness and multi-machine local swarms
Kimi K3 β evolutionary trading swarms today, local weights landing July 27
Grok / Grok Build β bound-folder agent loops users are asking to extend beyond git repos
Hermes Agent β the autonomous worker in manager-worker workflow engines
Codex β routed into other harnesses as the reasoning half of hybrid loops
OpenClaw β the default home for personal autonomous loops again this run (autonomous OSS contributor, scheduled morning briefs, telegram-connected home agents)
Claude Code β harness of choice for SEO delta crons, proxied GPT reasoning, and the reference design for DIY harnesses
CrewAI β the framework under both the viral open-source coding harness and multi-machine local swarms
Kimi K3 β evolutionary trading swarms today, local weights landing July 27
Grok / Grok Build β bound-folder agent loops users are asking to extend beyond git repos
Hermes Agent β the autonomous worker in manager-worker workflow engines
Codex β routed into other harnesses as the reasoning half of hybrid loops
Comments