Loop Daily: 2026-06-12
If yesterday was about pointing Fable at your hardest task, today the conversation moved up a level: not the task, the loop that runs the task. The single most-quoted idea on the timeline is that you should stop prompting agents and start designing the loop that prompts them, and the people actually doing it have stopped talking philosophy and started shipping verifiers. The strongest cases here all answer the same question in different ways: what does the loop check its own work against. A threat-detection agent grades itself on confidence, a CX agent on a benchmark it keeps sharpening, a security OS on tests it is not allowed to touch. The other thread, unavoidable, is cost. A loop re-sends its whole history every turn, so the difference between a smart loop and a runaway bill is the verification gate and the iteration cap.
#1
@HackingDave
https://x.com/HackingDave/status/2064821193006252256
This is the strongest production self-improving loop of the day, and it is not a demo. Binary Defense runs an agent called Scout Forge that looks over every customer submission, every log source and alarm, and asks one question on repeat: can we get better, with better training data, better normalization, new detection criteria. When a technology source it has never seen comes in, it automatically researches it, generates synthetic training data, and builds new parsers over time. The number that sells it: a new PLC source started at 13% confidence, hit 73% within 18 minutes, and reached 100% by end of day. Self-healing, self-improving threat detection that quietly gets better on its own.
https://x.com/HackingDave/status/2064821193006252256
This is the strongest production self-improving loop of the day, and it is not a demo. Binary Defense runs an agent called Scout Forge that looks over every customer submission, every log source and alarm, and asks one question on repeat: can we get better, with better training data, better normalization, new detection criteria. When a technology source it has never seen comes in, it automatically researches it, generates synthetic training data, and builds new parsers over time. The number that sells it: a new PLC source started at 13% confidence, hit 73% within 18 minutes, and reached 100% by end of day. Self-healing, self-improving threat detection that quietly gets better on its own.
#2
@OpenCovenant
https://x.com/OpenCovenant/status/2064636027340222838
Covenant is an agent-native OS where an autonomous loop writes, tests and ships its own code around the clock with every commit public, and this week it crossed from building to improving. They pointed it at one of its own core components, the engine that verifies its tamper-evident audit log, and let it rewrite the code it runs on. Eight rounds later it was 4x more efficient, better than they managed by hand, having taught itself vectorization and rewritten the underlying cryptography correctly along the way. The honest part is the guardrail: it cannot cheat, every rewrite has to produce identical results against tests it cannot touch or it gets rejected automatically. Recursive self-improvement is only scary as a black box, and the whole point here is verifiability.
https://x.com/OpenCovenant/status/2064636027340222838
Covenant is an agent-native OS where an autonomous loop writes, tests and ships its own code around the clock with every commit public, and this week it crossed from building to improving. They pointed it at one of its own core components, the engine that verifies its tamper-evident audit log, and let it rewrite the code it runs on. Eight rounds later it was 4x more efficient, better than they managed by hand, having taught itself vectorization and rewritten the underlying cryptography correctly along the way. The honest part is the guardrail: it cannot cheat, every rewrite has to produce identical results against tests it cannot touch or it gets rejected automatically. Recursive self-improvement is only scary as a black box, and the whole point here is verifiability.
#3
@AshwinSreenivas
https://x.com/AshwinSreenivas/status/2064759689381109774
Decagon shipped Duet Autopilot as a self-improving agent for customer experience and, smartly, built a benchmark to back the claim instead of just asserting it. DuetBench grades CX agents that learn over time, comparing Autopilot against certified human agent-builders on both outcome and methodology across 90 diagnostic investigations. What stands out is the behavior: rather than one-pass solving, Autopilot ran simulations, found broken branches, repaired the underlying tool, and repeated until the workflow passed. The headline result is the self-critique loop, it improved the quality of its own test set, lifting simulation accuracy from 58% to 88% across 520 runs. As these systems spread, verified evaluation starts to matter as much as raw model capability.
https://x.com/AshwinSreenivas/status/2064759689381109774
Decagon shipped Duet Autopilot as a self-improving agent for customer experience and, smartly, built a benchmark to back the claim instead of just asserting it. DuetBench grades CX agents that learn over time, comparing Autopilot against certified human agent-builders on both outcome and methodology across 90 diagnostic investigations. What stands out is the behavior: rather than one-pass solving, Autopilot ran simulations, found broken branches, repaired the underlying tool, and repeated until the workflow passed. The headline result is the self-critique loop, it improved the quality of its own test set, lifting simulation accuracy from 58% to 88% across 520 runs. As these systems spread, verified evaluation starts to matter as much as raw model capability.
#4
@shannholmberg
https://x.com/shannholmberg/status/2064700139235844220
The clearest piece of loop methodology on the timeline, on why coding loops and marketing loops are built differently. A coding loop has a hard signal to push against: tests pass, build passes, benchmark improves, the bug is gone, green means done. Marketing gets none of that, a weak landing page still loads, a bland post still publishes, nothing in the environment stops it. So a marketing loop needs judgment before autonomy, gates that act like tests for things a compiler can never check: truth, proof, specificity, voice, differentiation, taste, and a 'do nothing, the original is better' option. His sharp line: a coding loop ends at tests passed, a marketing loop should end at this is worth a human decision. Most AI marketing agents copy the motion of coding loops without the verification layer, so they produce more work, faster, with less taste.
https://x.com/shannholmberg/status/2064700139235844220
The clearest piece of loop methodology on the timeline, on why coding loops and marketing loops are built differently. A coding loop has a hard signal to push against: tests pass, build passes, benchmark improves, the bug is gone, green means done. Marketing gets none of that, a weak landing page still loads, a bland post still publishes, nothing in the environment stops it. So a marketing loop needs judgment before autonomy, gates that act like tests for things a compiler can never check: truth, proof, specificity, voice, differentiation, taste, and a 'do nothing, the original is better' option. His sharp line: a coding loop ends at tests passed, a marketing loop should end at this is worth a human decision. Most AI marketing agents copy the motion of coding loops without the verification layer, so they produce more work, faster, with less taste.
#5
@DeRonin_
https://x.com/DeRonin_/status/2064784790940008645
A concrete marketing loop that does have a self-grading step. He tested Higgsfield MCP, pasted one product URL, and got back a full set of creatives, videos, ads and a landing page, shipped straight to his directory with no separate UI and no prompt-engineering in another tab. The flow connects Higgsfield to Claude, Cursor, Perplexity or Hermes, then the agent plans the campaign, generates the assets, grades its own output, iterates, and ships. The whole creative stack in one agent loop. It is exactly the pattern the marketing-loop methodology calls for, the agent does the motion and a verification step decides what is good enough to ship.
https://x.com/DeRonin_/status/2064784790940008645
A concrete marketing loop that does have a self-grading step. He tested Higgsfield MCP, pasted one product URL, and got back a full set of creatives, videos, ads and a landing page, shipped straight to his directory with no separate UI and no prompt-engineering in another tab. The flow connects Higgsfield to Claude, Cursor, Perplexity or Hermes, then the agent plans the campaign, generates the assets, grades its own output, iterates, and ships. The whole creative stack in one agent loop. It is exactly the pattern the marketing-loop methodology calls for, the agent does the motion and a verification step decides what is good enough to ship.
#6
@WeixianXu
https://x.com/WeixianXu/status/2064529448213565831
A new autoresearch framework worth flagging: EEVEE, billed as the first multi-dataset test-time prompt learning framework for self-improving LLM agents. The framing matters, it is not a single-benchmark prompt-optimization story but built for agents facing the messy, shifting mixture of tasks real agents hit in the wild. The reported numbers: +42 cumulative improvement as tasks are added, +25% relative gain on Qwen3-4B-Instruct, and +61% relative on DeepSeek-V3.2. This is the academic underside of all the 'self-improving agent' talk, an actual method for keeping an agent getting better across heterogeneous real-world workloads after deployment.
https://x.com/WeixianXu/status/2064529448213565831
A new autoresearch framework worth flagging: EEVEE, billed as the first multi-dataset test-time prompt learning framework for self-improving LLM agents. The framing matters, it is not a single-benchmark prompt-optimization story but built for agents facing the messy, shifting mixture of tasks real agents hit in the wild. The reported numbers: +42 cumulative improvement as tasks are added, +25% relative gain on Qwen3-4B-Instruct, and +61% relative on DeepSeek-V3.2. This is the academic underside of all the 'self-improving agent' talk, an actual method for keeping an agent getting better across heterogeneous real-world workloads after deployment.
#7
@EnoReyes
https://x.com/EnoReyes/status/2064766716794872066
A clean three-line recipe for frontier AI research that doubles as a definition of autoresearch in practice: use models that support open research, run a mission in the desktop app with the goal of building out the component of your pipeline you care about, then monitor the agents for the duration, anywhere from two hours to two weeks. He calls it a GUI for auto research. The interesting part is the time horizon, this is not a one-shot prompt, it is a supervised long-run where the human role is to watch a process unfold over days, not to type the next step.
https://x.com/EnoReyes/status/2064766716794872066
A clean three-line recipe for frontier AI research that doubles as a definition of autoresearch in practice: use models that support open research, run a mission in the desktop app with the goal of building out the component of your pipeline you care about, then monitor the agents for the duration, anywhere from two hours to two weeks. He calls it a GUI for auto research. The interesting part is the time horizon, this is not a one-shot prompt, it is a supervised long-run where the human role is to watch a process unfold over days, not to type the next step.
#8
@kevintpayne
https://x.com/kevintpayne/status/2064608499359691126
A working example of agents running unsupervised for hours, self-improving as they go, with Fable 5 on Hyperagent. The two test cases are the tell: an asteroid visualization built from NASA data, and an Apollo control-panel reconstruction from PDFs. He frames these not as demos but as the kind of complex, multi-step work that only holds together when an agent can reason visually and course-correct on its own across a long run. The jump he points at is the real one for this whole category, from prompt an agent to do a task, to give it a goal and let it iterate until done.
https://x.com/kevintpayne/status/2064608499359691126
A working example of agents running unsupervised for hours, self-improving as they go, with Fable 5 on Hyperagent. The two test cases are the tell: an asteroid visualization built from NASA data, and an Apollo control-panel reconstruction from PDFs. He frames these not as demos but as the kind of complex, multi-step work that only holds together when an agent can reason visually and course-correct on its own across a long run. The jump he points at is the real one for this whole category, from prompt an agent to do a task, to give it a goal and let it iterate until done.
#9
@getsmallai
https://x.com/getsmallai/status/2064543242876850340
A practitioner shipped Small Harness 0.7.0, his first release built with Fable, and it is the observability release for agent loops. Two pieces matter: a flight recorder that drops an events JSONL sidecar for every session with tool calls, approvals, compaction and timing, and live nested subagent and critic activity; and an eval CLI that runs a bundled task end-to-end and exits 0/1, plus integration tests that drive the real agent loop against a mock SSE server with no live LLM needed. This is the unglamorous plumbing the whole loop conversation depends on, you cannot design a loop you cannot trace or test, and he open-sourced it.
https://x.com/getsmallai/status/2064543242876850340
A practitioner shipped Small Harness 0.7.0, his first release built with Fable, and it is the observability release for agent loops. Two pieces matter: a flight recorder that drops an events JSONL sidecar for every session with tool calls, approvals, compaction and timing, and live nested subagent and critic activity; and an eval CLI that runs a bundled task end-to-end and exits 0/1, plus integration tests that drive the real agent loop against a mock SSE server with no live LLM needed. This is the unglamorous plumbing the whole loop conversation depends on, you cannot design a loop you cannot trace or test, and he open-sourced it.
#10
@victorialslocum
https://x.com/victorialslocum/status/2064617082600272142
The most useful map of the agent-runtime chaos right now, breaking down OpenClaw, Hermes, Odysseus and n8n by what actually differs. The dimension that matters is persistent memory and self-improvement: most tools are session-based and forget everything when you close the window, which she calls one of the biggest bottlenecks in AI systems. Hermes was built to fix this by keeping memory across sessions and writing its own skill files from experience, OpenClaw has persistent memory via plugins you configure yourself, and in her hands Hermes is clearly better at the self-improvement loop, though both still have room. OpenClaw and Hermes are converging as autonomous, local-first, persistent runtimes, while Odysseus is a UI layer and n8n a low-code automation platform, two different categories entirely.
https://x.com/victorialslocum/status/2064617082600272142
The most useful map of the agent-runtime chaos right now, breaking down OpenClaw, Hermes, Odysseus and n8n by what actually differs. The dimension that matters is persistent memory and self-improvement: most tools are session-based and forget everything when you close the window, which she calls one of the biggest bottlenecks in AI systems. Hermes was built to fix this by keeping memory across sessions and writing its own skill files from experience, OpenClaw has persistent memory via plugins you configure yourself, and in her hands Hermes is clearly better at the self-improvement loop, though both still have room. OpenClaw and Hermes are converging as autonomous, local-first, persistent runtimes, while Odysseus is a UI layer and n8n a low-code automation platform, two different categories entirely.
#11
@jimboot
https://x.com/jimboot/status/2064620466371957019
The most concrete token-economics breakdown of an agentic loop today, and it ties the whole Loop and Super User conversation together. He traces one session: about 30 tool calls (web scrape, floorplan screenshots, file writes, QA screenshots), each re-sending the whole conversation, context grown to 130K+ tokens with the claude-api skill docs alone at ~70K, cumulative input across all requests probably 1.5 to 2M tokens. Caching saves it, Claude Code caches aggressively so most of that is cache reads at $1/M instead of $10/M, landing around $5-6 instead of the $18-20 it would have cost raw. The stealth cost was the screenshots, each image re-read on every request until cached. This is exactly the math behind every runaway-bill horror story this week.
https://x.com/jimboot/status/2064620466371957019
The most concrete token-economics breakdown of an agentic loop today, and it ties the whole Loop and Super User conversation together. He traces one session: about 30 tool calls (web scrape, floorplan screenshots, file writes, QA screenshots), each re-sending the whole conversation, context grown to 130K+ tokens with the claude-api skill docs alone at ~70K, cumulative input across all requests probably 1.5 to 2M tokens. Caching saves it, Claude Code caches aggressively so most of that is cache reads at $1/M instead of $10/M, landing around $5-6 instead of the $18-20 it would have cost raw. The stealth cost was the screenshots, each image re-read on every request until cached. This is exactly the math behind every runaway-bill horror story this week.
#12
@lividprowess
https://x.com/lividprowess/status/2064581324133007802
A 22-year-old biotech undergrad skipped learning boilerplate syntax and used a multi-agent loop to build an 11,000-line PyTorch framework simulating biophysical neuron populations, STDP, and BCI decoders. It is a small post but a real one, computational neuroscience is exactly the kind of non-software domain where the agent-loop pattern earns its keep, someone with the domain knowledge but not the coding reps using a loop to produce a serious research artifact. This is the same domain-knowledge-beats-syntax shift showing up in the Super User hackathon stories, just pointed at neurons instead of landing pages.
https://x.com/lividprowess/status/2064581324133007802
A 22-year-old biotech undergrad skipped learning boilerplate syntax and used a multi-agent loop to build an 11,000-line PyTorch framework simulating biophysical neuron populations, STDP, and BCI decoders. It is a small post but a real one, computational neuroscience is exactly the kind of non-software domain where the agent-loop pattern earns its keep, someone with the domain knowledge but not the coding reps using a loop to produce a serious research artifact. This is the same domain-knowledge-beats-syntax shift showing up in the Super User hackathon stories, just pointed at neurons instead of landing pages.
#13
@DavidShulmanFL
https://x.com/DavidShulmanFL/status/2064843547698721026
A clean personal application of Karpathy's autoresearch pattern: an unattended LLM loop with a program file that sets the objectives, runs to depth, and files output straight into a knowledge base. The twist is the input, instead of compiling the web it compiles his own conversations, the positions he holds, decisions he has made, projects, people, concepts. This is the autoresearch idea pointed inward, using the same unattended-loop-with-a-program-file structure not to discover new science but to build a structured, self-maintaining memory of one person's thinking.
https://x.com/DavidShulmanFL/status/2064843547698721026
A clean personal application of Karpathy's autoresearch pattern: an unattended LLM loop with a program file that sets the objectives, runs to depth, and files output straight into a knowledge base. The twist is the input, instead of compiling the web it compiles his own conversations, the positions he holds, decisions he has made, projects, people, concepts. This is the autoresearch idea pointed inward, using the same unattended-loop-with-a-program-file structure not to discover new science but to build a structured, self-maintaining memory of one person's thinking.
#14
@spenserskates
https://x.com/spenserskates/status/2064759773292368141
Amplitude's CEO introduced Wave, a proactive product agent that runs the whole build-ship-use-learn loop instead of just the building half. The argument is sharp: AI made building and shipping extraordinarily fast, but understanding usage and learning still happen by hand. Wave analyzes Amplitude data across analytics, feedback, session replays, error logs, agent traces and experiment results, surfaces opportunities as full product specs you or your agents can approve and ship, then tracks the outcomes so the loop starts over. It is a product launch, but it is a genuine attempt to close the learning half of the loop that most agent setups leave to humans.
https://x.com/spenserskates/status/2064759773292368141
Amplitude's CEO introduced Wave, a proactive product agent that runs the whole build-ship-use-learn loop instead of just the building half. The argument is sharp: AI made building and shipping extraordinarily fast, but understanding usage and learning still happen by hand. Wave analyzes Amplitude data across analytics, feedback, session replays, error logs, agent traces and experiment results, surfaces opportunities as full product specs you or your agents can approve and ship, then tracks the outcomes so the loop starts over. It is a product launch, but it is a genuine attempt to close the learning half of the loop that most agent setups leave to humans.
#15
@victor_zhng
https://x.com/victor_zhng/status/2064812227228823817
A compact, useful methodology for building a corporate-finance agent loop, from someone actually doing finance. Assuming you already have the harness (prompt instructions, context management, retrieval, agent loop, access control, tool calling), the method is: break down the process into retrieval, financial modeling and output production, break down the capacities you want to judge such as retrieval accuracy, calculation accuracy, reasoning and judgment calls, design the evals accordingly, then run, find problems, tune the harness and test again. It is the eval-driven harness-tuning discipline applied to a high-stakes non-coding domain, and it is the unglamorous version of everything the loop-engineering crowd is gesturing at.
https://x.com/victor_zhng/status/2064812227228823817
A compact, useful methodology for building a corporate-finance agent loop, from someone actually doing finance. Assuming you already have the harness (prompt instructions, context management, retrieval, agent loop, access control, tool calling), the method is: break down the process into retrieval, financial modeling and output production, break down the capacities you want to judge such as retrieval accuracy, calculation accuracy, reasoning and judgment calls, design the evals accordingly, then run, find problems, tune the harness and test again. It is the eval-driven harness-tuning discipline applied to a high-stakes non-coding domain, and it is the unglamorous version of everything the loop-engineering crowd is gesturing at.
π‘ Eco Products Radar
Eco Products Radar
Fable (11 mentions) is the model everyone is wiring into their loops this cycle. Hermes (8) and OpenClaw (3) remain the two persistent, local-first agent runtimes the self-improvement conversation centers on, with Cursor (5) and Claude Code (5) as the coding harnesses. MCP (6) is the connective layer for tool-using loops. EEVEE (4) shows up as the academic self-improving-agent framework of the day, and DeepSeek (3) recurs as both a benchmark target and a cheap model for the cost-conscious loop builders.
Fable (11 mentions) is the model everyone is wiring into their loops this cycle. Hermes (8) and OpenClaw (3) remain the two persistent, local-first agent runtimes the self-improvement conversation centers on, with Cursor (5) and Claude Code (5) as the coding harnesses. MCP (6) is the connective layer for tool-using loops. EEVEE (4) shows up as the academic self-improving-agent framework of the day, and DeepSeek (3) recurs as both a benchmark target and a cheap model for the cost-conscious loop builders.
Comments