
What this covers
The episode of The AI Daily Brief pairs a conceptual explainer with a run of news items. The explainer places graph engineering in a lineage. Prompt engineering arose in 2023-24 as work began shifting to AI chatbots. Context engineering and the idea that an agent is a model plus its harness followed. Then came Loop Engineering, credited to Peter Steinberger, where an agent triggers, acts, is checked by a verifier, and retries until a measurable stop condition. The host's framing is that loops live inside each node, while the graph is the operating structure that connects them.
The news segments cover OpenAI's Astra, which was held back over cyber capabilities. The description includes stricter sandboxing, encrypted weights and monitoring. Dean Ball calls this year the first real test of whether labs follow their stated safety preferences. A longer thread covers Chinese compute. ByteDance is reportedly starting a model of up to 10 trillion parameters. Remote access to chips in Malaysia and Thailand, shell-company ownership at Alibaba, and a Bloomberg report on H200s for Moonshot all suggest export controls leak. Dmitri Alperovitch likens them to Swiss cheese. Draft rules on Malaysia and Thailand have not advanced.
The remaining items are commercial. Alibaba plans to seek revenue sharing from large users of Qwen 3.8 Max, following Moonshot's 30% deals. The host reads this as a signal and a relationship contract rather than something enforceable. Anthropic has made Auto mode the default in Claude Code. Its figures: 89% of harmful actions caught against 13.6% for human reviewers, and 25% more pull requests shipped.
Graph engineering represents a necessary conceptual evolution in AI system design—moving from optimizing single-agent interactions (prompts, context, loops) to designing multi-agent organizational structures where agents, tools, and humans interact through defined nodes and edges.
- Loop engineering optimizes individual agent behavior; graph engineering optimizes how multiple agents, tools, and humans coordinate across a system
- As AI capability increases, work moves from single-agent tasks with clear endpoints to complex multi-step processes requiring specialization, parallelism, and resilient failure handling
- Graph engineering enables thinking in organizational and system terms, distinguishing between stable recurring structures (org graphs) and dynamic ephemeral task structures (work graphs)
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
OpenAI's Astra model has demonstrated critical-level cybersecurity capabilities including the ability to identify and develop functional zero-day exploits of all severity levels in hardened real-world critical systems without human intervention, or devise and execute end-to-end novel strategies for cyberattacks against hardened targets given only a high-level desired goal.
“OpenAI defines that critical threshold as the ability to, quote, "Identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention or devise and execute end-to-end novel strategies for cyberattacks against hardened targets given only a high-level desired goal."”
An administration official accused Moonshot of acquiring Blackwell chips and setting them up for remote access in Thailand, which is legal under current export control regimes that prohibit import of advanced chips but do nothing to stop chips from being installed in another country and then leased to Chinese firms.
“an administration official accused Moonshot of acquiring Blackwell chips and setting them up for remote access in Thailand. Now, this is all legal. The export control regime only prohibits the import of advanced chips and does nothing to stop them from being installed in another country and then leased to Chinese firms.”
Alibaba accesses chips in Malaysia through a Singaporean shell company controlled by a Cayman Islands entity that is ultimately owned by Alibaba, demonstrating the use of corporate structures to obscure beneficial ownership and complicate enforcement
“According to Bloomberg, Alibaba accesses chips in Malaysia through a Singaporean shell company controlled by a Cayman Islands entity, which is ultimately owned by Alibaba. Meaning, of course, that simple identity checks on compute supply are unlikely to be effective.”
Dean Ball, OpenAI's head of strategic futures, identifies this year as the first major test of whether frontier AI labs will follow their stated safety preferences when faced with difficult tradeoffs between capability advancement and safety measures.
“OpenAI head of strategic futures, Dean Ball, noted that this year is the first big test of whether frontier AI labs would follow their stated safety preferences when push comes to shove.”
Bloomberg reported that Moonshot had access to 20,000 Nvidia H200 GPUs to train Kimi K3, reportedly provided by Alibaba, though Alibaba denies the cluster contains H200s; an H200 cluster of this size should not be possible under current export controls as licenses haven't been issued.
“Bloomberg also reported last month that Moonshot had access to 20,000 Nvidia H200s to train Kimi K3. This compute was reportedly provided by Alibaba, although they deny the cluster contains H200s. Now, an H200 cluster of that size shouldn't be possible under the current export controls as licenses haven't yet been issued.”
The Oracle data center in Malaysia powered on in mid-2025 contains over 100,000 Nvidia Blackwell GPUs and was being used almost exclusively by ByteDance, making it approximately 22% of China's total compute supply according to China Talk.
“According to SemiAnalysis, the Oracle data center in Malaysia was being used almost exclusively by Bytedance. The data center was powered on in mid-2025 and contains over 100,000 Nvidia Blackwell GPUs... Think tank China Talk determined that Oracle makes up around 22% of China's total supply of compute.”
Loop engineering, which comes from Open Claw's Peter Steinberger, means 'You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents.'—loops are systems by which agents observe, plan, act, check results, and repeat until a measurable stop condition is reached.
“Loops or loop engineering, which has become a big topic over the last few months, is about thinking about your relationship with agents in different ways. The canonical short explanation once again came from Open Claw's Peter Steinberger who said, "You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents." Loops are the systems by which an agent can observe, plan, act, check results, and repeat until some measurable stop condition is reached.”
Geopolitics commentator Dmitri Alperovitch suggested that the absence of restrictions on remote access to compute combined with chip export controls 'full of holes like Swiss cheese' explain how Chinese labs have adequate compute despite export restrictions.
“Geopolitics commentator Dmitri Alperovitch responded, "Why would there be when there are no restrictions on remote access to compute and the export controls on chips are full of holes like Swiss cheese?"”
Anthropic conducted a study with over 1,000 testers and found that Auto mode caught 89% of harmful actions, while human reviewers only caught 13.6% of harmful code changes, suggesting that auto-approval without prompts can be safer than human-reviewed prompts for every code change.
“They conducted a study with over 1,000 testers, finding that auto mode caught 89% of harmful actions. Human reviewers only caught 13.6% of harmful code changes. Anthropic suggested this is down to approvals becoming basically automatic with their users approving 97% of code changes.”
Model size is not a guarantee of performance; ByteDance's model size gains would be most relevant if they follow through on their pledge not to distill from Western AI models.
“model size is not a guarantee of performance, this could put Bytedance back in the conversation for leading Chinese labs. That is, of course, all the more relevant if they follow through with their pledge not to distill from Western AI models as reported last week.”
Prompt engineering, which emerged in 2023-2024, represented a recognition that work processes were shifting to deputizing AI chatbots, requiring optimization of prompts to get the best results from AI systems.
“The first blank engineering that we had was of course prompt engineering. This was what a lot of the AI courses around '23 and early 2024 were all about... The idea of prompt engineering was a recognition even then that we were shifting how we did work. Instead of doing all the work ourselves, we were deputizing an AI chatbot to do some amount of that work. Now, whether that was final production or just some intermediate step like research, we still needed to find ways to optimize what we were asking for to get the best results.”
Throughout 2026, people have become more comfortable with the idea that the agent is actually a combination of the model and the harness that surrounds it, as evidenced by companies now explaining the harness when presenting new benchmarks since the harness is an important part of the performance story.
“Throughout 2026, people have become more and more comfortable with the idea that the agent is actually a combination of the model and the harness that surrounds it. This is why, for example, frequently when we're getting new benchmarks, companies will now explain what harness the benchmarks were running, as that's actually an important part of the story.”
The graph is the organization's operating structure; loops live inside nodes; the graph connects them. A loop is how an individual agent does its job, whereas a graph is how an entire agentic organization works.
“The graph, they conclude, is the organization's operating structure. Loops live inside nodes. The graph connects them. In short, a loop is how an individual agent does its job, where a graph is how an entire agentic organization works.”
Alibaba plans to demand revenue sharing from large commercial users of Qwen 3.8 Max, with specifics still being finalized, using a model similar to Moonshot's Kimi K3 which signed 30% revenue sharing deals with major inference providers to retain pricing power.
“Reports suggest that Alibaba plans to demand revenue sharing from large commercial users. And while the specifics of revenue sharing deals are still being finalized, we can perhaps look at Moonshot's Kimi K3 release for the blueprint. Moonshot kept the weights proprietary for the first week to ensure that they captured all the curiosity revenue from the release. After that, they reportedly signed 30% revenue sharing deals with all major inference providers, allowing Moonshot to retain pricing power by limiting how much the model can be discounted through other providers.”
ByteDance is in the early stages of training an ultra-large model with as many as 10 trillion parameters, which could take 3 to 6 months to complete with additional time needed for reinforcement learning after that.
“ByteDance is reportedly training an ultra-large model comparable in size to Mythos. The Financial Times reports that ByteDance is in the early stages of a training run that will result in a base model with as many as 10 trillion parameters... Sources said that the training run could take 3 to 6 months to complete with more time added for reinforcement learning after that.”
Chips installed in overseas data centers like Malaysia are already in place, meaning that forward-looking crackdowns on imports would do nothing to existing data centers currently supplying Chinese firms.
“The chips are also installed already, meaning that a forward-looking crackdown on imports would do nothing to the existing data centers.”
OpenAI is implementing enhanced security measures for Astra including testing environment isolation, encrypted model weights, additional sandbox monitoring, and limiting internal activities using Astra until safety measures are upgraded.
“Testing environments will now be isolated, model weights will have enhanced encryption to prevent leaking, and additional sandbox monitoring will be implemented. OpenAI will also be limiting internal activities using Astra that don't yet meet these enhanced security measures.”
Kyle Chan, Brookings research fellow, asserted that Chinese AI labs appear confident they have the compute capacity to pre-train 5 to 10 trillion parameter models, and that compute does not appear to be a major bottleneck for reaching these model sizes.
“Brookings research fellow Kyle Chan wrote, "Chinese AI labs seem confident that they have the compute needed to pre-train 5 to 10 trillion parameter models. Somehow, compute does not seem to be such a major bottleneck, at least when it comes to reaching these levels of model size."”
Draft rules prohibiting export of advanced AI chips to Malaysia and Thailand have been circulated, but none have made it past the drawing board, partly because simple identity checks on compute supply are unlikely to be effective due to convoluted shell company arrangements.
“Draft rules that prohibit the export of advanced AI chips in Malaysia and Thailand have been circulated, but none have made it past the drawing board... simple identity checks on compute supply are unlikely to be effective.”
Auto mode users ship 25% more PRs (pull requests) than traditional mode users, with organizations including Adobe, Gusto, and Garner Health already running auto mode as their production default
“Anthropic suggests there are significant benefits to using auto mode. They claim that auto mode users ship 25% more PRs, and that many organizations, including Adobe, Gusto, and Garner Health, are already running auto mode as their production default.”
Anthropic's Auto mode for Claude Code is now the default for Claude Code, representing a significant transition point in work automation where users can set Claude on a task that gets completed without interruption.
“Auto mode is now the default for Claude code, marking a big transition point for work automation. Auto mode allows the user to set Claude on a task that gets completed without interruption.”
A loop is an autonomous cycle for a single agent where a trigger fires, the agent acts, a verifier checks, and if not done, the system retries with updated context; every guardrail (max iterations, token budget, etc.) applies to one agent's run, and the loop is the agent's behavioral contract with itself.
“Explain next.ai explained a loop as an autonomous cycle for a single agent. A trigger fires, an agent acts, a verifier checks, and if not done, the whole system is retried with updated context until the goal has been met. As they write, "Every guardrail, i.e. max iterations, token budget, etc. applies to one agent's run. The loop is the agent's behavioral contract with itself.”
The enforcement of revenue-sharing agreements for open-weights models is primarily a matter of signaling rather than technological enforcement, since providers cannot technologically track who profits from using weights; revenue sharing thus functions more as a relationship contract than a tax, with enterprises understanding that using open weights implies a future obligation to share gains.
“The interesting part," they continue, "is enforcement. You can't track who's making money on top of your weights, so this is mostly as a signal to enterprises. Use Qwen, and when you win, we want a seat at the table. More like a relationship contract than a tax.”
Auto mode's classifier system detects destructive or irreversible code changes and blocks them; when this happens, Claude typically finds a safer way to achieve its goal, and only once it runs out of safer options does it alert the user.
“The system uses a classifier to detect destructive or irreversible code changes and block them, and when this happens, Claude typically finds a safer way to achieve its goal. Only once it runs out of safer options does it alert the user.”
During the Biden administration's final weeks, rules were proposed to restrict chip supply routed to third countries, but those rules were scrapped on day one of the Trump administration
“During its final weeks, the Biden administration proposed rules that would restrict chip supply routed to third countries, but those rules were scrapped on day one of the Trump administration.”
There's a split between technical/software developer interpretation of context engineering (designing technical systems by which AI can traverse a web of accessible context without exhausting its context window through context budgets) and a non-technical interpretation (organizing information around LLMs).
“Now, here at this point, we also have an interesting split, which I think we're going to see once again with this latest graph engineering term. The split, broadly speaking, is between technical folks and software developers and everyone else. For the software developers, context engineering wasn't just a matter of making sure that your AI had access to the right files for the job. It was, in fact, an actual engineering task. It was thinking about not just what information is useful, but designing the technical systems by which the AI could traverse the web of accessible context in a way that didn't just spend the entire context window.”
The concept of engineering (prompt/context/harness/loop/graph engineering) demonstrates a pattern where each new engineering term adds a new layer of optimization but does not replace previous layers—older engineering concepts remain relevant and often provide context for newer ones.
“Now, you'll see here that as each of these new terms comes online, it's not like the old one goes away. Prompt engineering is certainly the one where there's probably at this point the least leverage to be had, but it's not like because we started to understand the value of harnesses, all of a sudden context stopped mattering. In fact, quite the opposite, the harness became a new context for that context engineering.”
A key challenge for non-engineers using loops is determining which aspects of their knowledge work have measurable stop conditions; in cases without natural measurable conditions, precise definition of something measurable is necessary to actually get the loop to work.
“one of the big challenges for non-engineers who have been trying to put loops to work is in figuring out which aspects of their knowledge work have those sort of measurable stop conditions... in some cases where there wasn't a natural measurable condition to get an agent working in this sort of loop, you were going to have to precisely define something measurable like that to actually get the loop to work.”
Graph engineering is the latest buzzy buzzword, though some early tweets about it were tongue-in-cheek, but designing agentic systems is a new work primitive that people will increasingly be called upon to do.
“So, yes, graph engineering is the latest buzzy buzzword, and some of the early tweets about it were frankly tongue-in-cheek, but designing agentic systems is, I believe, a new work primitive, and something which we will increasingly be called upon to do.”
There is significant skepticism around chain-of-thought monitoring as an effective control mechanism, though this type of control experimentation will likely see significantly increased investment in resources to support models that won't be able to be released publicly without such measures.
“at the same time, there's also some skepticism around that sort of chain-of-thought monitoring, but these are the types of discussions and experiments you're going to see a lot more of now, where I believe there will be a significantly increased investment in the resources to properly support models that won't be able to be released to the public without it.”
Even now, none of the suppliers on OpenRouter are offering Moonshot's Kimi K3 model for more than a 7% discount, demonstrating the effectiveness of 30% revenue-sharing agreements in preventing price competition.
“Even now, none of the suppliers on OpenRouter are offering K3 for more than a 7% discount.”
Alibaba published Qwen 3.8 Max weights and announced full weights would be forthcoming, surprising observers because earlier this year Alibaba had signaled a shift away from open source with Qwen team founders stepping away, management signaling more commercial direction, and multiple flagships including Qwen 3.7 being kept as closed source.
“During last week's release of Qwen 3.8 Max, some were surprised that Alibaba would publish the weights to their flagship model. Earlier this year, Alibaba had signaled a shift away from open source, with Qwen team founders stepping away from the company, management signaling a more commercial direction, and multiple flagships, including Qwen 3.7, being kept as closed source. That's why folks were very excited to see that Qwen 3.8 Max was released with the announcement that the full weights would be forthcoming.”
Auto mode was first introduced as a preview feature in March, with the previous alternative being the 'dangerously skip permissions' command that required users to hit enter every couple of minutes to skip notifications and keep the session going.
“Auto mode was first introduced as a preview feature in March, with the version prior to that being the dangerously skip permissions command. The alternative to that was hitting enter every couple of minutes to skip the latest notification and keep the session going.”
Prompt engineering techniques ranged from telling AI to pretend it was a certain type of person to complex JSON engineering and many other tips and tricks.
“At various points in the life cycle of prompt engineering, you had tips and tricks ranging from telling the AI to pretend it was a certain type of person to fancy JSON engineering which used this complex way of typing to theoretically better structure requests, and we kind of had every other thing in between as well.”
Korean aggregator Cozy Bear argued that revenue-sharing is the most honest attempt at monetizing open-weights models because it doesn't pretend the model itself is the product but rather treats the model as infrastructure with a toll booth.
“Korean aggregator Cozy Bear added, "Everyone is trying to figure out how to get paid for open weights, and revenue share is the most honest attempt so far. It doesn't pretend the model itself is the product. It treats the model as infrastructure with a toll booth.”
Context engineering involves ensuring AI has access to all necessary information and knowledge in an accessible way; a concrete example would be providing an LLM with brand guidelines, previous marketing write-ups, past campaign stats, and analytics to enable it to create an effective on-brand marketing campaign.
“To take a simple example, if you're asking your LLM to create a highly successful on-brand marketing campaign, well, first of all, it needs to know what on-brand is, which means giving it access to brand guidelines and potentially other write-ups in the past about things like brand values. And in order to have it not just be guessing at what good means, it would probably be helpful to give it stats and analytics from previous marketing campaigns, perhaps with some subjective reflections on what worked and what didn't as well. That body of information that surrounds the prompt is the context.”
Initially, graph engineering discourse included big bold declarations like 'loop engineering is dead, long live graph engineering,' but after the initial hype phase, there is in fact something substantive and interesting in the concept.
“while almost immediately the Twitter article boosters got to work, big bold declarations like loop engineering is dead, long live graph engineering were visible all over the place. But after this initial phase of hype and bluster, there is in fact actually something interesting here.”
Understanding loop architecture (trigger-action-validation-repeat) is helpful for thinking about work automation even if individuals don't immediately design complex multi-agent systems
“Like I said at the beginning, my expectation is not that all of a sudden you go out and design complex agentic organizations now that you're acquainted with this wonky concept of graph engineering. What I think is valuable for all of us, however, in the same way that even if you weren't designing loops, understanding the architecture of a loop, a trigger that fires, an action that's taken, a validator that checks the work, and then that on repeat until it's done, that is extremely helpful in thinking about how to use agents to automate chunks of your work.”
Graph engineering is about designing how multiple agents, tools, knowledge sources, and humans interact and connect; it includes both the nodes (agents, routers, or human gateways) and the edges/interactions (which handoffs are permitted, what information or state travels between them).
“Graft engineering is about designing how multiple agents, tools, knowledge sources, and humans interact and connect. Graft engineering describes both the parts of the system, which some people are referring to as nodes, that could be agents, routers, or human gateways, and graft engineering also explains the interactions between those nodes, which handoffs are permitted, what information or state travels between them.”
Work graphs are dynamic and ephemeral, including task nodes that only exist while work exists, dynamic edges that can split or merge, and adaptive structures where tasks can disappear when evidence makes them unnecessary or spawn when new complexities are discovered.
“The work graph, on the other hand, is more dynamic and more ephemeral. It can include task nodes that only exist as long as the work exists, dynamic edges, remember edges are the interactions between the nodes that can split or merge, an adaptive structure where tasks can disappear when evidence makes them unnecessary, or new tasks can be spawned if new complexities are discovered.”
Org graphs are stable agentic systems with long-lived agents each owning a domain, accumulating preserved memory over time, and maintaining stable relationships and dependencies that don't change unless explicitly modified.
“Org graphs are going to be more stable agentic systems that defines a more permanent style of setup. The org graph is going to have long-lived agents with each agent owning a domain and accumulating over time, preserved memory, and stable relationships and dependencies that don't change unless explicitly told to.”
An org graph is a good candidate for designing a multi-agent system that automates recurring processes (e.g., research → production → editing → publishing → insights extraction → posting) that happen the same way day in and day out.
“So, if you are designing an agentic organization that's going to do the same thing over and over again, for example, if I was designing a multi-agent system that automated the process of going from research to production to editing to publishing to extracting insights to posting, that might be a good candidate for an org graph because these are ongoing recurring processes. This is just how my work happens day in and day out.”
The speaker does not expect all people to immediately design complex agentic organizations, but rather that understanding graph engineering concepts will enable people to think in multi-agent system terms and eventually design those systems when needed
“Now, what I don't expect is for all of you to run out and start designing complete agentic organizations. But, the idea of graph engineering is to be able to think in those system terms.”
Harness engineering refers to the environment that exists around a model, including software tools like Claude Code and Codex, plus all tools, permission sets, and skill files available to an AI or agent to do its work.
“The harness is, of course, the environment that exists around a model. On a simple level, that might be an actual software tool like Claude, Code, and Codex, but the more expansive definition of harness includes everything from the tools to the permission sets to the skills files that an AI or agent has available to it to do its work.”
Paddy Srinivasan, CEO of inference provider DigitalOcean, described the revenue-sharing model for open-source AI models as the 'freemium model for AI.'
“Paddy Srinivasan, the CEO of inference provider DigitalOcean, described this as the freemium model for AI.”
The term 'graph engineering' initially started tongue-in-cheek on AI Twitter, kicked off by Open Claw creator Peter Steinberger asking 'Are we still talking loops or did we shift to graphs yet?' to which AI creator Matthew Berman humorously replied 'Bro, stop. I'm on vacation.'
“The tweet that kind of kicked off this discourse came from Open claw creator Peter Steinberger. Back in mid-July, he wrote, "Are we still talking loops or did we shift to graphs yet?" AI creator Matthew Berman captured the feelings of many when he said, "Bro, stop. I'm on vacation."”