Loop Daily: August 15, 2026
Two things collided today. DeepSeek shipped a harness where the agent loop itself is a swappable plugin — and where an agent can inspect its own runtime, write a new plugin, mount it, and use that capability in the next task. Meanwhile the loop skeptics landed a real punch: a Thoughtworks evaluation put TDD inside the agent loop and found it produced worse code at higher token cost. The connective tissue running through everything else is a unit-of-measure argument. Cost per run is dead, cost per successful task is the number, and on OpenRouter agentic traffic now consumes 13 to 15 times more tokens per request than human traffic. The loop is no longer a curiosity — it is the dominant consumer of inference, and people are starting to price it accordingly.
#1
@bennyjo
https://x.com/bennyjo/status/2087781529241969117
The cleanest running example of a self-improving loop in public. Phil is an AI agent that learns by paper-trading Polymarket prediction markets, then applies what it learns to real markets through Pearl Connect. Every hour it researches, then trades or skips, then edits its own playbook based on what settled. Before every trade it writes down its estimate, its reasoning and the edge it claims, and that entry is frozen in a ledger it cannot edit. Every night it runs a deep retrospective across everything that settled and audits its own decisions the way a stranger would. Today's update is that it derived its most impactful rule to date from its own losses, wrote it into its own playbook, and has obeyed it since — six for six against the market.
https://x.com/bennyjo/status/2087781529241969117
The cleanest running example of a self-improving loop in public. Phil is an AI agent that learns by paper-trading Polymarket prediction markets, then applies what it learns to real markets through Pearl Connect. Every hour it researches, then trades or skips, then edits its own playbook based on what settled. Before every trade it writes down its estimate, its reasoning and the edge it claims, and that entry is frozen in a ledger it cannot edit. Every night it runs a deep retrospective across everything that settled and audits its own decisions the way a stranger would. Today's update is that it derived its most impactful rule to date from its own losses, wrote it into its own playbook, and has obeyed it since — six for six against the market.
#2
@CoreyGallon
https://x.com/CoreyGallon/status/2087711598378975360
Richard Socher opened "First Steps Toward Automated AI Research" with a joke that functions as the thesis: every time I fire a linguist, my accuracy goes up, so maybe it is time to fire the AI engineers and have them manage an AI that does AI research instead. The concrete result is the part worth reading. On NanoChat, a benchmark the community had already been grinding on, the auto-research system moved 0.93 to 0.91 bits per byte in a day or two — and not by tuning hyperparameters. It found hashed bigram and trigram embedding tables mixed into attention value paths through learned gates. That is an architectural discovery, not a search over a config space, which is the distinction that determines whether auto-research is real.
https://x.com/CoreyGallon/status/2087711598378975360
Richard Socher opened "First Steps Toward Automated AI Research" with a joke that functions as the thesis: every time I fire a linguist, my accuracy goes up, so maybe it is time to fire the AI engineers and have them manage an AI that does AI research instead. The concrete result is the part worth reading. On NanoChat, a benchmark the community had already been grinding on, the auto-research system moved 0.93 to 0.91 bits per byte in a day or two — and not by tuning hyperparameters. It found hashed bigram and trigram embedding tables mixed into attention value paths through learned gates. That is an architectural discovery, not a search over a config space, which is the distinction that determines whether auto-research is real.
#3
@allenholub
https://x.com/allenholub/status/2087698126119330166
The most useful negative result of the day. He notes that his own process — he specifies the tests, then tells the AI to make them pass — works well. Then he read Birgitta Böckeler's write-up from Thoughtworks on putting TDD inside the agent loop: instruct the AI to write a test and then make it pass, with no human in between. The conclusion was not merely that it failed to help, but that it made the code worse and cost more in tokens. His explanation is the interesting bit: TDD is a discipline about how humans work. Writing the test first keeps a person focused on what the code needs to do; writing only enough code to pass keeps the system minimal. None of those pressures apply to a model.
https://x.com/allenholub/status/2087698126119330166
The most useful negative result of the day. He notes that his own process — he specifies the tests, then tells the AI to make them pass — works well. Then he read Birgitta Böckeler's write-up from Thoughtworks on putting TDD inside the agent loop: instruct the AI to write a test and then make it pass, with no human in between. The conclusion was not merely that it failed to help, but that it made the code worse and cost more in tokens. His explanation is the interesting bit: TDD is a discipline about how humans work. Writing the test first keeps a person focused on what the code needs to do; writing only enough code to pass keeps the system minimal. None of those pressures apply to a model.
#4
@tokenbytoken_ai
https://x.com/tokenbytoken_ai/status/2087853847070839008
The same Thoughtworks evaluation, with the numbers stated plainly: no clear benefit from TDD inside the loop, non-TDD runs often scored higher on design, and they used roughly a third of the tokens. Two independent accounts of the same finding on the same day is about as close as this field gets to replication, and it points at a broader trap — human process discipline does not automatically survive being handed to an agent.
https://x.com/tokenbytoken_ai/status/2087853847070839008
The same Thoughtworks evaluation, with the numbers stated plainly: no clear benefit from TDD inside the loop, non-TDD runs often scored higher on design, and they used roughly a third of the tokens. Two independent accounts of the same finding on the same day is about as close as this field gets to replication, and it points at a broader trap — human process discipline does not automatically survive being handed to an agent.
#5
@jiayuan_jy
https://x.com/jiayuan_jy/status/2087911060154314963
He was pulled into the DeepSeek Harness repo a month ago when it was a shell with only the core framework implemented, and says every pull since has brought thousands of commits. His reading: DSH is simultaneously a runnable coding agent and an agent development framework. The core claim is that everything is a plugin — model, tools, filesystem, shell, sandbox, session storage, subagent, UI, and the agent loop itself. The part that matters for this newsletter is what that unlocks: DSH can already have an agent inspect its own runtime, write a plugin on the spot, mount it, and then use that newly acquired capability in subsequent tasks. He is honest that it is experimental — dynamically generated plugins live only in memory and vanish on restart, with no path yet to settling into a permanent plugin.
https://x.com/jiayuan_jy/status/2087911060154314963
He was pulled into the DeepSeek Harness repo a month ago when it was a shell with only the core framework implemented, and says every pull since has brought thousands of commits. His reading: DSH is simultaneously a runnable coding agent and an agent development framework. The core claim is that everything is a plugin — model, tools, filesystem, shell, sandbox, session storage, subagent, UI, and the agent loop itself. The part that matters for this newsletter is what that unlocks: DSH can already have an agent inspect its own runtime, write a plugin on the spot, mount it, and then use that newly acquired capability in subsequent tasks. He is honest that it is experimental — dynamically generated plugins live only in memory and vanish on restart, with no path yet to settling into a permanent plugin.
#6
@mylifcc
https://x.com/mylifcc/status/2087896239455310322
A genuinely deep comparison of DSH against Pi. Pi aims at a minimal core that users assemble themselves; DSH plugin-izes almost everything and ships several complete official combinations. The skeleton is Cordis, which is not an ordinary plugin loader — it enforces reversible side effects, meaning that when a plugin unmounts, its registered services, events, tool schemas and prompt sections are all automatically withdrawn, plus dependency declaration where plugins declare what services they need via inject. The session design is the hardcore part: sessions are append-only event streams, and the context the model finally sees must be fully reconstructable from that log. The official rule is stated bluntly as model-visible means logged. Fork, resume, replay, UI rendering and telemetry all project from the same stream, which he argues is more thorough than most coding agents manage.
https://x.com/mylifcc/status/2087896239455310322
A genuinely deep comparison of DSH against Pi. Pi aims at a minimal core that users assemble themselves; DSH plugin-izes almost everything and ships several complete official combinations. The skeleton is Cordis, which is not an ordinary plugin loader — it enforces reversible side effects, meaning that when a plugin unmounts, its registered services, events, tool schemas and prompt sections are all automatically withdrawn, plus dependency declaration where plugins declare what services they need via inject. The session design is the hardcore part: sessions are append-only event streams, and the context the model finally sees must be fully reconstructable from that log. The official rule is stated bluntly as model-visible means logged. Fork, resume, replay, UI rendering and telemetry all project from the same stream, which he argues is more thorough than most coding agents manage.
#7
@MinLiBuilds
https://x.com/MinLiBuilds/status/2087891094465925196
His framing is that DeepSeek did not open-source a Claude Code clone, it open-sourced its entire opinion about how agents should be built — model, tools, filesystem, shell, session, skill, subagent, workflow, context management, security policy, all decomposed into swappable modules. Think of it as an agent breadboard where even the loop is replaceable. The detail he finds funniest is that most people shipping a pure engineering project would write a README and call it done, whereas DeepSeek attached a paper, "A Programming Paradigm for Spatiotemporal Composability," explaining the theory behind the plugin system.
https://x.com/MinLiBuilds/status/2087891094465925196
His framing is that DeepSeek did not open-source a Claude Code clone, it open-sourced its entire opinion about how agents should be built — model, tools, filesystem, shell, session, skill, subagent, workflow, context management, security policy, all decomposed into swappable modules. Think of it as an agent breadboard where even the loop is replaceable. The detail he finds funniest is that most people shipping a pure engineering project would write a README and call it done, whereas DeepSeek attached a paper, "A Programming Paradigm for Spatiotemporal Composability," explaining the theory behind the plugin system.
#8
@SahilPanhotra
https://x.com/SahilPanhotra/status/2087899089438724398
The compressed version for anyone who just wants the shape: DeepSeek launched its own agent harness, and the interesting part is that almost every piece can be swapped — model, tools, skills, sandbox, filesystem, agent loop, orchestration — under MIT. An agent loop being a configuration choice rather than a framework property is a structural shift, not a feature.
https://x.com/SahilPanhotra/status/2087899089438724398
The compressed version for anyone who just wants the shape: DeepSeek launched its own agent harness, and the interesting part is that almost every piece can be swapped — model, tools, skills, sandbox, filesystem, agent loop, orchestration — under MIT. An agent loop being a configuration choice rather than a framework property is a structural shift, not a feature.
#9
@DecurityHQ
https://x.com/DecurityHQ/status/2087831083563855934
A tightly scoped agent loop doing something a general assistant cannot. verifyoor takes an unverified Solidity smart contract that exists only as EVM runtime bytecode on chain and loops until it has reconstructed source plus compiler settings that recompile to byte-identical bytecode. The termination condition is objective and binary — the bytes match or they do not — which is exactly the property that makes an autonomous loop trustworthy. No human judgment needed in the middle, and demo plus source are both public.
https://x.com/DecurityHQ/status/2087831083563855934
A tightly scoped agent loop doing something a general assistant cannot. verifyoor takes an unverified Solidity smart contract that exists only as EVM runtime bytecode on chain and loops until it has reconstructed source plus compiler settings that recompile to byte-identical bytecode. The termination condition is objective and binary — the bytes match or they do not — which is exactly the property that makes an autonomous loop trustworthy. No human judgment needed in the middle, and demo plus source are both public.
#10
@tom_doerr
https://x.com/tom_doerr/status/2087771020409143448
Ralph is a minimal file-based agent loop for autonomous coding that treats files and git as memory. The whole design thesis is in that sentence: no vector database, no bespoke memory service, just the durable artifacts a repository already produces. Commit history becomes the trace, the working tree becomes the state, and the loop restarts from disk rather than from context.
https://x.com/tom_doerr/status/2087771020409143448
Ralph is a minimal file-based agent loop for autonomous coding that treats files and git as memory. The whole design thesis is in that sentence: no vector database, no bespoke memory service, just the durable artifacts a repository already produces. Commit history becomes the trace, the working tree becomes the state, and the loop restarts from disk rather than from context.
#11
@anna_y_zhang
https://x.com/anna_y_zhang/status/2087731800118530306
The sharpest diagnosis of why most self-improvement efforts stall. Agent traces are the most important new input to a self-improving company, but a pile of traces improves nothing by itself. The full loop needs three components: a company brain holding shared memory of decisions, context and the why behind the work; the traces themselves, meaning the raw record of what every human and agent tried and what got thrown away; and a learning loop that converts those traces back into better context, better skills and better next runs. Most teams have some of the first and are starting to accumulate the second. Almost nobody has the third, and only the ones who close it compound.
https://x.com/anna_y_zhang/status/2087731800118530306
The sharpest diagnosis of why most self-improvement efforts stall. Agent traces are the most important new input to a self-improving company, but a pile of traces improves nothing by itself. The full loop needs three components: a company brain holding shared memory of decisions, context and the why behind the work; the traces themselves, meaning the raw record of what every human and agent tried and what got thrown away; and a learning loop that converts those traces back into better context, better skills and better next runs. Most teams have some of the first and are starting to accumulate the second. Almost nobody has the third, and only the ones who close it compound.
#12
@stretchcloud
https://x.com/stretchcloud/status/2087829512080052382
The chart nobody is reading carefully enough. On OpenRouter, agentic token consumption crossed above human consumption on February 6, 2026. Since then total platform tokens went from roughly 500 billion to 7.3 trillion on a seven-day average as of August 10. Human usage grew 2.8x to 1.4 trillion; the remaining 5.9 trillion is agentic, growing orders of magnitude faster. The ratio that matters: agentic requests consume 13 to 15 times more tokens per request than human ones. His conclusion is that current API pricing, capacity planning and rate limit design were all calibrated for a human-session pattern that is no longer the dominant mode, and that cost per token is simply the wrong unit.
https://x.com/stretchcloud/status/2087829512080052382
The chart nobody is reading carefully enough. On OpenRouter, agentic token consumption crossed above human consumption on February 6, 2026. Since then total platform tokens went from roughly 500 billion to 7.3 trillion on a seven-day average as of August 10. Human usage grew 2.8x to 1.4 trillion; the remaining 5.9 trillion is agentic, growing orders of magnitude faster. The ratio that matters: agentic requests consume 13 to 15 times more tokens per request than human ones. His conclusion is that current API pricing, capacity planning and rate limit design were all calibrated for a human-session pattern that is no longer the dominant mode, and that cost per token is simply the wrong unit.
#13
@JinseokKim60030
https://x.com/JinseokKim60030/status/2087789051378085915
One sentence that explains most failed self-improvement loops: the hard part is not routing, it is the metric. A self-improving loop only converges if the scorer is harder to game than the task, otherwise the agent optimizes your eval rather than your goal. Cheap models make the loop affordable; a trustworthy grader makes it useful. Anyone building an overnight optimization harness should treat this as the design constraint rather than an afterthought.
https://x.com/JinseokKim60030/status/2087789051378085915
One sentence that explains most failed self-improvement loops: the hard part is not routing, it is the metric. A self-improving loop only converges if the scorer is harder to game than the task, otherwise the agent optimizes your eval rather than your goal. Cheap models make the loop affordable; a trustworthy grader makes it useful. Anyone building an overnight optimization harness should treat this as the design constraint rather than an afterthought.
#14
@JinseokKim60030
https://x.com/JinseokKim60030/status/2087790911610659133
The companion argument on economics. Cost per task matters more than cost per run, because a cheaper model that needs one extra correction round on a long agent loop erases a 2x price gap almost immediately. He suggests pairing any price claim with tokens spent and retry count before calling it a win. This is the discipline that most model-comparison threads skip entirely.
https://x.com/JinseokKim60030/status/2087790911610659133
The companion argument on economics. Cost per task matters more than cost per run, because a cheaper model that needs one extra correction round on a long agent loop erases a 2x price gap almost immediately. He suggests pairing any price claim with tokens spent and retry count before calling it a win. This is the discipline that most model-comparison threads skip entirely.
#15
@tickswiz
https://x.com/tickswiz/status/2087887918631141697
The counterpoint, with numbers. DeepSeek made V4-Pro official at 87.9 on Terminal-Bench against Fable 5's 88, at a listed $0.44 per million input and $0.88 output versus Fable's $10 and $50. His framing: the 5 percent capability gap is a slide, the 50x price gap is the product. If your agent loop burns tokens for forty minutes at a stretch, you are not really buying intelligence, you are buying a rate card. Read alongside the cost-per-task argument above, these two form the actual debate.
https://x.com/tickswiz/status/2087887918631141697
The counterpoint, with numbers. DeepSeek made V4-Pro official at 87.9 on Terminal-Bench against Fable 5's 88, at a listed $0.44 per million input and $0.88 output versus Fable's $10 and $50. His framing: the 5 percent capability gap is a slide, the 50x price gap is the product. If your agent loop burns tokens for forty minutes at a stretch, you are not really buying intelligence, you are buying a rate card. Read alongside the cost-per-task argument above, these two form the actual debate.
#16
@devcansado404
https://x.com/devcansado404/status/2087871452099826018
A pointed warning about flexible reasoning parameters. They look clean until an agentic loop escalates to max effort on a malformed payload and burns the token budget in minutes. His conclusion is that runtime governance needs hard guardrails, otherwise max reasoning is just unbounded cost and latency wearing a feature's clothes. Every autonomous loop with a dynamic effort dial needs a ceiling that the loop itself cannot raise.
https://x.com/devcansado404/status/2087871452099826018
A pointed warning about flexible reasoning parameters. They look clean until an agentic loop escalates to max effort on a malformed payload and burns the token budget in minutes. His conclusion is that runtime governance needs hard guardrails, otherwise max reasoning is just unbounded cost and latency wearing a feature's clothes. Every autonomous loop with a dynamic effort dial needs a ceiling that the loop itself cannot raise.
#17
@Montyclt
https://x.com/Montyclt/status/2087820650165551227
A production agent loop in customer support, described by someone who implemented it. When a ticket opens the model does nearly everything: analyzes it, routes it to the correct department, diagnoses it, proposes a solution to the human agent, and gives that agent a button that lets the model apply the solution through an agent loop with tool calling, then drafts the customer reply. For phone calls a Whisper-class model transcribes in real time while an LLM proposes solutions as the conversation happens. This is the shape autonomous loops take when they land in a real operations org — human at the approval gate, agent doing the traverse.
https://x.com/Montyclt/status/2087820650165551227
A production agent loop in customer support, described by someone who implemented it. When a ticket opens the model does nearly everything: analyzes it, routes it to the correct department, diagnoses it, proposes a solution to the human agent, and gives that agent a button that lets the model apply the solution through an agent loop with tool calling, then drafts the customer reply. For phone calls a Whisper-class model transcribes in real time while an LLM proposes solutions as the conversation happens. This is the shape autonomous loops take when they land in a real operations org — human at the approval gate, agent doing the traverse.
#18
@BestAIToolFind
https://x.com/BestAIToolFind/status/2087737367209882066
Prime Agent, from Prime Intellect, is an open-source self-improving coding harness built around a Recursive Language Model abstraction. It launched on Product Hunt on August 10, 2026 and finished sixth for the day with 174 upvotes. It sits alongside their open reinforcement-learning training stack — prime-rl, verifiers, Environments Hub — and a GPU marketplace, which makes it one of the few self-improving harnesses shipped by a group that also owns the training loop underneath it.
https://x.com/BestAIToolFind/status/2087737367209882066
Prime Agent, from Prime Intellect, is an open-source self-improving coding harness built around a Recursive Language Model abstraction. It launched on Product Hunt on August 10, 2026 and finished sixth for the day with 174 upvotes. It sits alongside their open reinforcement-learning training stack — prime-rl, verifiers, Environments Hub — and a GPU marketplace, which makes it one of the few self-improving harnesses shipped by a group that also owns the training loop underneath it.
#19
@SciFi
https://x.com/SciFi/status/2087908614975443405
A paper worth the title alone: "Agentic Auto-Research is Fuzz Testing," from Yifeng He, Jicheng Wang, Yinzhe Zhao, Jiachen Liu and Hao Chen. The reframing is useful — if you treat auto-research as fuzzing rather than as reasoning, then coverage, mutation strategy and crash detection become the design vocabulary, and the value comes from volume of cheap trials against a reliable oracle rather than from any single clever step.
https://x.com/SciFi/status/2087908614975443405
A paper worth the title alone: "Agentic Auto-Research is Fuzz Testing," from Yifeng He, Jicheng Wang, Yinzhe Zhao, Jiachen Liu and Hao Chen. The reframing is useful — if you treat auto-research as fuzzing rather than as reasoning, then coverage, mutation strategy and crash detection become the design vocabulary, and the value comes from volume of cheap trials against a reliable oracle rather than from any single clever step.
#20
@Secondmindsys
https://x.com/Secondmindsys/status/2087760855601402231
Why small accuracy gains stop being incremental once loops get long. On a one-shot task, five percent better feels marginal. Across a long agent loop, fewer errors means less recovery, less rework, and more of the task surviving to completion. He is careful not to call it AGI, but the structural point stands: reliability compounds across the length of the work, so better models unlock disproportionately more useful agents rather than proportionately better ones.
https://x.com/Secondmindsys/status/2087760855601402231
Why small accuracy gains stop being incremental once loops get long. On a one-shot task, five percent better feels marginal. Across a long agent loop, fewer errors means less recovery, less rework, and more of the task surviving to completion. He is careful not to call it AGI, but the structural point stands: reliability compounds across the length of the work, so better models unlock disproportionately more useful agents rather than proportionately better ones.
#21
@discernmentlab
https://x.com/discernmentlab/status/2087742728834830585
Three signals in one week that the second brain is becoming agent-queryable rather than human-browsable: Notion acquired ZeroEntropy for AI retrieval, Genspark launched SecondBrain and named the problem AI amnesia, and Oracle published a self-improving second brain guide. The shift he names is from storage to self-maintaining knowledge — the memory layer stops being a place you put things and becomes something that runs its own maintenance loop.
https://x.com/discernmentlab/status/2087742728834830585
Three signals in one week that the second brain is becoming agent-queryable rather than human-browsable: Notion acquired ZeroEntropy for AI retrieval, Genspark launched SecondBrain and named the problem AI amnesia, and Oracle published a self-improving second brain guide. The shift he names is from storage to self-maintaining knowledge — the memory layer stops being a place you put things and becomes something that runs its own maintenance loop.
#22
@mr_bailando
https://x.com/mr_bailando/status/2087880077883081151
Jack Dorsey shipped a free, open-source AI agent OS at 26.2k GitHub stars and still climbing, framed as infrastructure for running entire businesses through agents with OS-level abstraction and full agentic loop support. His read is a business one: most agent infra startups charging $1,000-plus a month are selling orchestration, persistent memory and task routing, and this commoditizes all three at once, for free, with real distribution behind it. The moat was never the orchestration.
https://x.com/mr_bailando/status/2087880077883081151
Jack Dorsey shipped a free, open-source AI agent OS at 26.2k GitHub stars and still climbing, framed as infrastructure for running entire businesses through agents with OS-level abstraction and full agentic loop support. His read is a business one: most agent infra startups charging $1,000-plus a month are selling orchestration, persistent memory and task routing, and this commoditizes all three at once, for free, with real distribution behind it. The moat was never the orchestration.
#23
@cortensor
https://x.com/cortensor/status/2087705634313797814
An honest mid-build devlog rather than a launch post. PyClaw is heading into an end-of-month assessment reviewing how far the agent loop, coding workflows, memory, tooling and overall usability have actually progressed, with the explicit goal of identifying what still blocks a practical MVP. They already expect an API compatibility gap against Cortensor Portal, and plan to use PyClaw itself to surface the remaining gaps — pointing the agent at its own integration surface is a small, real self-improvement loop.
https://x.com/cortensor/status/2087705634313797814
An honest mid-build devlog rather than a launch post. PyClaw is heading into an end-of-month assessment reviewing how far the agent loop, coding workflows, memory, tooling and overall usability have actually progressed, with the explicit goal of identifying what still blocks a practical MVP. They already expect an API compatibility gap against Cortensor Portal, and plan to use PyClaw itself to surface the remaining gaps — pointing the agent at its own integration surface is a small, real self-improvement loop.
#24
@VibeCoderOfek
https://x.com/VibeCoderOfek/status/2087912037242253465
The line that summarizes the day: cost per successful task is the only number that survives contact with an overnight agent loop, and everything else is just the leaderboard version.
https://x.com/VibeCoderOfek/status/2087912037242253465
The line that summarizes the day: cost per successful task is the only number that survives contact with an overnight agent loop, and everything else is just the leaderboard version.
📡 Eco Products Radar
Eco Products Radar
DeepSeek Harness (DSH) — the day's center of gravity, MIT-licensed, where model, tools, sandbox, filesystem, orchestration and the agent loop itself are all swappable plugins.
Cordis — the plugin kernel underneath DSH, notable for reversible side effects on unmount and dependency-injected spatial composability, shipped with its own paper.
Pi — the minimal-core counterpoint to DSH's batteries-included plugin philosophy, recurring as the comparison baseline.
Claude Code — still the default reference harness that every new loop framework positions against.
Codex — the other constant comparison point in harness benchmarking threads.
Polymarket — the venue of choice for public self-improving agent loops, where settlement provides a free, unfakeable grader.
Prime Agent — open-source self-improving coding harness built on a Recursive Language Model abstraction, from the team that also runs prime-rl and Environments Hub.
Ralph — minimal file-based agent loop that treats files and git as the memory layer instead of a vector store.
DeepSeek Harness (DSH) — the day's center of gravity, MIT-licensed, where model, tools, sandbox, filesystem, orchestration and the agent loop itself are all swappable plugins.
Cordis — the plugin kernel underneath DSH, notable for reversible side effects on unmount and dependency-injected spatial composability, shipped with its own paper.
Pi — the minimal-core counterpoint to DSH's batteries-included plugin philosophy, recurring as the comparison baseline.
Claude Code — still the default reference harness that every new loop framework positions against.
Codex — the other constant comparison point in harness benchmarking threads.
Polymarket — the venue of choice for public self-improving agent loops, where settlement provides a free, unfakeable grader.
Prime Agent — open-source self-improving coding harness built on a Recursive Language Model abstraction, from the team that also runs prime-rl and Environments Hub.
Ralph — minimal file-based agent loop that treats files and git as the memory layer instead of a vector store.
Comments