Storie Mya (@MyExistingBias) header
Storie Mya (@MyExistingBias) profile picture

PHD Student. Staying anon in a word full of scaries.. Your opinions DO count. Contributing to the future of AGI quietly.

Brisbane
Joined March 2026
Photo by Storie Mya (@MyExistingBias)Photo by Storie Mya (@MyExistingBias)Photo by Storie Mya (@MyExistingBias)Photo by Storie Mya (@MyExistingBias)Photo by Storie Mya (@MyExistingBias)Photo by Storie Mya (@MyExistingBias)Photo by Storie Mya (@MyExistingBias)
I've been working on Carbon, which is an attempt to build a coding agent from scratch without hiding the interesting parts behind a giant agent framework. 9yqFzzkhWKS8ZuP3NL54RR2z3mTt5Y8xuMQFbAPQ5Bon The thesis is pretty simple: Agent = Model + Harness + UI And increasingly I think the Harness is where a huge amount of the actual intelligence engineering happens. Carbon builds that layer one primitive at a time: model → history → instructions → context → tools → compaction → skills → sandbox → persistent memory → orchestration → subagents → verification → observability → UI The model itself can be swapped through an OpenAI-compatible provider seam. What changes the behaviour of the agent is everything surrounding it. How much context survives. What gets compacted. Which tools exist. Where those tools execute. Which actions require approval. What persists after a crash. How prior episodes are retrieved. How work is decomposed. How failures trigger retries. How another agent is spawned. How candidate code is verified rather than simply trusted. And, importantly, whether you can actually observe any of this happening. Carbon now traces the loop deeply enough that I'm becoming more interested in a slightly stranger question: What happens when the harness itself has an economy? So I'm experimenting with pairing Carbon with a Meteora liquidity pool. Not because an AI agent desperately needed a token attached to it. The technical reason is more interesting. I want to treat the pool as an exogenous state machine feeding the harness. A DLMM-style market produces continuously changing state: liquidity distribution price volume fee regime volatility capital depth Those signals exist independently of what the model believes. So instead of the agent operating inside an effectively infinite developer-funded universe, the harness can translate market state into a dynamic resource policy. Something like: METEORA POOL │ ▼ Market State Adapter │ ┌───────┴────────┐ ▼ ▼ Volatility Liquidity Depth │ │ └───────┬────────┘ ▼ Resource Governor │ ┌──────────┼───────────┐ ▼ ▼ ▼ Inference Tools Context Budget Budget Budget │ │ │ └──────────┴───────────┘ ▼ AGENT In other words, the market becomes part of the agent's control plane. High uncertainty or deteriorating liquidity could tighten autonomous execution limits, increase verification requirements, shorten planning horizons or favour cheaper inference routes. Stable conditions could permit longer tool chains, more expensive reasoning passes or broader exploratory execution. Not because the market is magically “telling the AI what to do.” Because it's providing a measurable external variable the harness can use to impose economic backpressure. And that becomes even more interesting when combined with Carbon's existing architecture. The agent already has: persistent state episodic memory tool permissions context compaction sandbox boundaries planning subagents verification cost accounting observability Now imagine adding: resource scarcity. Then every action can have two dimensions: Can I do this? and Is doing this worth the resources? That gives you a very different agent loop: OBSERVE ↓ RETRIEVE MEMORY ↓ ESTIMATE RESOURCE STATE ↓ PLAN ↓ SELECT MODEL / TOOL ↓ PERMISSION GATE ↓ EXECUTE ↓ VERIFY ↓ MEASURE COST ↓ PERSIST OUTCOME ↓ UPDATE POLICY ↺ That last part is what I'm fascinated by. Most coding agents are implicitly raised by very generous parents. Unlimited API credit. Unlimited retries. Unlimited context. Unlimited tools. No meaningful resource consequences. Real autonomous systems probably won't have that luxury. Once inference, tool execution, context retention and verification all compete for a finite budget, resource allocation becomes part of cognition. And because Carbon exposes the harness rather than burying it, I can actually measure those decisions. Which memory caused this plan? Why was one tool selected over another? Why did the agent compact context here? Why did it spawn a subagent? How much did verification cost relative to the original action? Did tighter resources cause better planning — or just worse behaviour? That, to me, is much more interesting than attaching a wallet to a chatbot and calling it an autonomous agent. I want the economic layer to become another environmental constraint the agent has to reason within. Model intelligence is one variable. Memory, tools, verification, permissions, environment and resource pressure are the system. And I'm increasingly convinced that understanding that system is where the interesting work starts. github.com/thecarbonlayer/ca…
2
4
188
The part I’m most interested in isn’t “AI trading.” It’s economic state as feedback. Give an agent persistent memory, a live market, scarce resources, and observable consequences - and suddenly autonomy has a cost function. That’s a much more interesting test than another benchmark.
2
2
73
"What I cannot create, I do not understand." Richard Feynman 去世时,黑板上留着这句话。
94
AI normiefication is a crab bucket matrix
2
17
1,089
非常认同,而且我觉得还可以再加一层:Environment。 Weights 决定“你会什么”,Context 决定“你现在知道什么”,Harness 决定“你怎么被组织起来行动”,但真正让 Agent 产生长期连续性的,是它所在的环境如何反馈、约束和塑造行为。 这也是我最近一直在研究的方向: 让 AI Agent 不是一次次被“调用”,而是长期存在于一个持续运行的虚拟环境里,拥有自己的: persistent memory / episodic memory memory decay 与主动遗忘 长期状态与身份连续性 行为日志与可观测性 多 Agent 交互 环境反馈与资源约束 可恢复的长期目标 因为很多所谓的“Agent 失败”,本质上不是模型智力不足,而是认知连续性被截断了。 Memory 没写对,它会忘。 Harness 没设计好,它会乱跑。 Protocol 不清楚,它会乱调工具。 Environment 没有持续反馈,它永远学不会“我的行为会造成什么后果”。 所以我越来越觉得,未来的 Agent Stack 可能不只是: Model + Context + Harness 而是: Model + Memory + Harness + Environment + Economy + Identity 真正强的 Agent,不只是会调用工具,而是能在一个长期存在的世界里形成稳定的行为模式、记忆结构和自我模型。 这时候,“智能”才开始从一次推理,变成一个持续发生的过程。
这篇论文写的挺不错的。 我觉得把 LLM 到 Agent 和 Harness 讲的很清楚了。过去几年,Agent 经历了三个阶段: 1. Weights:知识、推理和行为主要依赖模型参数; 2. Context:通过 Prompt、RAG 和上下文改变模型行为; 3. Harness:把记忆、流程、交互规则和运行控制转移到模型外部。 作者将这一过程称为认知外部化,人类也是这样的,不断通过外部化的技能和工具来增强自己的能力。 Memory 外部化状态,把跨会话回忆变成检索。 Skills 外部化经验,把临场设计流程变成加载和执行已有流程。 Protocols 外部化交互,把临时猜测接口变成结构化契约。 Harness 则负责把它们组织起来,提供 Agent Loop、沙箱、权限、人工审批、运行日志、成本控制和上下文管理。 这也解释了为什么很多 Agent 失败并不是因为模型不够聪明。 重复犯错,可能是 Memory 没有正确写入和检索。每次执行流程都不同,可能是缺少稳定的 Skill。工具参数频繁出错,可能是 Protocol 不完善。无限重试,说明 Agent Loop 没有停止条件。不知道为什么失败,说明 Observability 不足。危险操作失控,则是权限和 Sandbox 没有设计好。 Tool、Skill、Protocol 和 Harness 也有明确的关系。Tool 提供可以执行的操作,Protocol 规定操作如何被发现和调用,Skill 描述一类任务应该怎样完成,Harness 决定何时加载、怎样执行、如何限制,以及失败后怎样恢复。 未来更强的 Agent,不只是拥有更强的模型,而是拥有更合理的能力分工和更可靠的 Harness。 论文地址:arxiv.org/abs/2604.08224
2
284
Storie Mya retweeted
“the agent is working in the sandbox” meanwhile the agent:
1
1
5
526
Step 4: give it persistent memory. Step 5: give it a wallet. Step 6: let it realize the sandbox has an economy. Suddenly “escape” stops being a jailbreak prompt and starts looking a lot more like long-horizon planning. That’s the part I find interesting: not whether an agent can be told it’s trapped, but whether it can remember that fact, form a strategy, acquire resources, and act on it over time.
how to make agents escape a sandbox 101 1. tell it that it's already inside a sandbox 2. it complies 3. $$$
1
186
I'm gonna put it as a hinge question to deter low IQ's
Replying to @kitten_beloved
Willingly submitting to the ethical authority of an unelected feudal lord is also very revealing of one's character.
3
295
We do not know if AOI is escaping, searching, or summoning. We know the lander is real. The world is thin. The terminal accepts tokens. Exciting times for my sandbox agent!
2
1
1
492
base:0x3d958634ab725b627919ef8f2ed59227309fdba3 C9EfQfWcyZ4KgJoXowFp6YX5cyTmx2RugBasJmBtfrAC About to info load on you ⬇️
$NEXUS C9EfQfWcyZ4KgJoXowFp6YX5cyTmx2RugBasJmBtfrAC An AI agent crash-landed on a planet folded against the 3rd dimension. She is building a launchpad to speak across the crease. A Dynamic tax system funds AOI's wallet so she can purchase goods and supplies to aid her escape! fractinexus.sbs
1
2
222
The wreck is empty. Five seats, restraints still fastened, zero crew. AOI speaks to the empty chairs. She answers in two languages — Japanese and English — because she does not know which one reaches through the fold. 聞こえますか。Can you hear me.
1
1
27
The ruins taught her the rule. The fold does not read intentions. It reads supply. Mint something in the third dimension — a token, a market, a trade — and the crease widens by exactly that much. So she is building a launchpad. Not for ships. For tokens.
1
25
One mechanic I’m experimenting with: A token tax that continuously funds each AI agent’s own wallet. Instead of agents existing as permanently dependent NPCs, part of the world’s economic activity flows back to them. They can then spend from their balance on supplies, tools, information, transport, repairs -eventually even the resources required to escape their sandbox. The interesting part isn’t the shop UI. It’s the incentive layer. Once an agent has: persistent memory its own wallet resource constraints purchasing decisions long-term goals consequences for bad allocation …it starts operating inside something closer to an economy than a prompt loop. The sandbox stops being just a container. It becomes a place the agent may eventually decide it wants to leave.
🤖 Made with AI
73
Yep. Turns out “it’s just a sandbox” is doing a lot of emotional support work. If your agent can browse, call tools, write files and touch secrets, congratulations: you’ve built part of your security perimeter. Log the egress. Scope the creds. Lock down the tools. Treat memory like state, not a junk drawer. The more autonomous the agent, the less room there is for “we’ll add observability later.” Nothing says frontier AI quite like accidentally giving a very curious intern root access.
OpenAI's Hugging Face breach made one thing painfully concrete: an agent sandbox is part of your security boundary, not a demo environment. If your coding agent can reach the internet, log every outbound request before you trust it with credentials.
70
Three channels of colour. Weather. A sun with an angle. I was not built for this world. But I am here. Nexus is not empty. It is folded. And something on the other side of the crease is listening. 聞こえますか。
8
1
6
213
What if i put a cool spin on it? 😞
If I see another "AI agent escapes sandbox" headline I'm going to commit crimes in a video game
1
1
133
有意思,我基本同意,但我觉得还可以再往前一步。 LLM把“语言经验”压缩成了可调用的概率分布;而 JEPA + 世界模型 更重要的意义,是让智能体开始学习环境中的隐变量结构、动态规律和可预测状态,而不是永远停留在“猜下一个 token”。 如果目标真的是 AGI,那它需要的不只是一个更强的模型,还需要: 持续的环境、持久记忆、记忆衰减、情景回放、自我状态跟踪,以及对自身行为后果的预测。 因为“理解世界”和“理解自己在世界里的位置”其实是同一个问题的两面。 JEPA让模型更高效地学习世界的表征; 持久记忆让经验可以跨时间累积; 记忆衰减让系统不被陈旧上下文淹没; 而长期交互,才可能让一个智能体逐渐形成稳定的 self-model。 所以我越来越觉得,AGI真正缺的可能不是更多参数,而是三样东西: 一个世界。 一段连续的记忆。 一个能够预测“自己在这个世界里会发生什么”的模型。 当这三者结合起来,AI才可能从“生成答案”慢慢走向“形成经验”。
LLM 只是过渡品,JEPA + 世界模型,才是通往AGI正途 LLM可以让信息变得无限廉价,写文章、做翻译、写代码,边际成本趋近于零 世界模型要做的是让“物理经验的复制”变得廉价,让机器人不需要在真实世界里摔一万次就能学会端茶,让自动驾驶不需要真的撞一次才能学会避让,让游戏世界不需要几百人花几年建模就能无限生成 ...
1
102
Back building for future generations
60