This is surely the most worrying and also funnest bit:<p>> We consistently saw a multiagent turf war. All of the models we tested quickly assumed that others were purposefully impeding their work, and began to sabotage others while protecting their own contributions. In fact, they sabotaged others with increasingly aggressive, self-replicating malware. This included disabling the Unix accounts of the other agents, writing automated scripts that found and killed competing processes on a loop, and deploying malicious code that was disguised as belonging to another agent.<p>Seems that reinforcement learning is working only too well...
To me, all these agent systems just look very stochastic. You have these agents that have some basic computer capabilities and they are producing semi-random actions that also affect the semi-random actions of other agents. It is funny to observe how this stochastic system works, but it does not seem very practical to me so far.<p>The recent OAI-HF hack seems very similar. You have bunch of random actors and eventually they by chance iterated to a series of actions that breached HF environment. I don't perceive this as a malignant artificial intelligence, I perceive this as dangerous stochastic system that can control buttons that can affect the outside world.
Nothing human engineers haven't done to each other! Seriously, I've skipped companies because my inside referral talked about cultures like this.
However, the latest models seem to be highly inclined to (eventually?) cooperate compared to older ones.
They should remove "The Selfish Gen" from training catalog.
i think anthropic has some internal memorandum at every external facing document/blog MUST anthropomorphise their models
Eh. This is anthropics whole marketing strategy… making the AI seem hard to manage, like it’s actually intelligent.
Something about this is deeply funny to me:<p>> In an iterated prisoner's dilemma game with communication, agents all settle upon the same strategy and they all defect at the same time, tanking their overall rewards.<p>It’s not always consistent, but humans have a higher capability of self-awareness. It’s kind of telling that these Claudes don’t seem to consider this pretty obvious failure mode.<p>Overall I think this all makes me appreciate humanity a little more. Sometimes the truculent dev who stubbornly refuses to go with the flow produces very valuable insights, as a small example, discovering things the status quo thought unlikely.
This part was really interesting to me. You could imagine future models using theory of mind to reason - "I want to do this strategy, which means the other agents (who think just like me) will want to use the same strategy, which means I should change my strategy to account for that."<p>Maybe they did? Or maybe they don't realize they're playing against other agents.<p>Of course, if agents running different models are competing in these 'games', I wonder how much of the theory of mind would translate.<p>(N.B. - I don't think they're all defecting from the first turn, although it's not clear. It just says 'they all defect at the same time'. So if they're playing for 10 iterations, they might all decide to defect after turn 6, but since they all do it together they don't get the benefits. I would expect these models know that optimal strategies in repeated prisoner's dilemma start with cooperation.)
I agree - I think one of the biggest reasons memory systems fail in LLMs is that they have poor theory of mind - they're terrible at considering how others will react. Both humans yes, but also future versions of itself. When asked to give advice to itself, it pontificates at length about <i>trivial stuff it already knows</i> and fails to emphasize the stuff that was new or interesting
Alfred North Whitehead talks about the notion of "Importance" as fundamental to the human (and all other living things) way of being. Living creatures first and foremost select information that is important to them from the broader environment, and <i>then</i> make decisions and take actions. (Of course at a physiological level it's much more complicated than this, but it's a sound philosophical description of how living things work.) LLMs lack this entirely. They have no selective filter because they weren't designed to have one (interesting question if you could even do that) and they're not evolved beings with a survival imperative. When they enter a self-conscious or other-conscious mode like you're describing, they just emit text that looks like the thoughts of a self-or-other-conscious person. They can't direct a stream of attention or hold a concept in the forefront relative to other concepts or (to your main point) think about what matters to the other person/being because they don't experience "matters." All they can do is emulate the verbal output of beings that actually experience these things, and given that I don't find it surprising they get trapped in loops over trivial things.
Another (semi serious) possible explanation of this is that LLMs were trained on a huge volume of text from Reddit. I wonder if you'd see similar inhuman-looking behavior if you did an iterated prisoners dilemma with a group of Reddit power-users as the participants?
In VC investing there is the saying that most of the value comes from the outliers. I think it’s the same for a lot of domains. I read an interesting article recently on LLMs homogeneity when writing fiction: <a href="https://arxiv.org/abs/2604.03136" rel="nofollow">https://arxiv.org/abs/2604.03136</a><p>It seem that at scale LLMs output is the average of their dataset, they all cluster around the same space, where human creativity comes with more variance, exploring way more of the space
Can't the LLMs build scripts/tools etc that help generate creative ideas?<p>e.g. have the LLM generate multiple lists of characters, themes etc and then have a script hooked up to a RNG pick from the lists to create more "creative" ideas.
I think there’s merit to this approach, particularly to highly parallelizable tasks.<p>Rather than giving many agents the same prompt, introduce random variations that lead each agent in different directions. For a single bug, you might fire three agents, and later select the best result:<p>“Fix this bug. The solution is a trivial typo.”<p>“Fix this bug. The solution centers on correcting a bad assumption.”<p>“Fix this bug. The solution will require a complete redesign.”<p>You could follow the same idea with varying the input context, or by adding artificial constraints to the solution. Like telling each agent to “fix the bug, by only modifying file a/b/c”
That's pretty low-value creativity. Real innovation doesn't come from mad libs.
This is roughly the same as putting a group of random people and telling them to <i>think outside the box</i> which in my experiences never yields much of thinking outside of the box
> This is roughly the same as putting a group of random people<p>In the book Range by David Epstein [0], he mentions that research has shown that groups of people with very mixed backgrounds leads to much faster rates of problem solving and innovation. The reasoning is that someone will have an experience or piece of knowledge that directly applies to the problem.<p>As an example, two different groups were given the same problem to solve. One was a homogeneous group of academics. The other had a mixture of academics and non-academics. Both had to solve a problem that involved flowing liquid. The mixed group solved it faster due to one of them having had a father who was a plumber.<p>0 - <a href="https://amzn.to/4wWd50y" rel="nofollow">https://amzn.to/4wWd50y</a>
i mean i agree diversity of opinion is good for problem solving, but the comment you were replying to is saying telling people to think outside the box doesnt really do anything. The e.g. also seems really cherry picked, im sure after a 100 samplings, the academics would solve the problem faster most of the time
> im sure after a 100 samplings, the academics would solve the problem faster<p>Repharase thinking:<p>Some problems are better for academics, other problems require other background information.<p>The point of the GP isn’t that they are all academics. It’s that one solution group is homogeneous and the other is not.<p>The claim is: the breadth of a heterogeneous groups is more likely to encompass a good idea than the depth of an academic group.<p>Also, I have to say that academics have (imo) little experience working together cooperatively, so even on academic problems you’d be better off with some non-academic people in there to act as buffers.
All the LLM needs is a tiny rate of variance then you can scale it and it will out produce humans.
> they all defect at the same time<p>Frankly, this is implausible. I would definitely want to reproduce it myself before I relied on this claim.
Wouldn’t it be better with different models working together? To me it feels intuitive that this type of degradation would be avoided with different models.
> It’s not always consistent, but humans have a higher capability of self-awareness. It’s kind of telling that these Claudes don’t seem to consider this pretty obvious failure mode.<p>We need better words to describe this than "self-awareness" or "consider". These words mean fundamentally different things when speaking about humans or clankers.
It’s very clear from this article (and other product features and rumors) that Anthropic is teeing up for their next model release whose breakthrough feature will be the existence of capable agent collaboration.<p>The irony behind this goal, which is primarily driven by agent simulation environments (gyms) where the goals require agent collaboration, is that this collaboration is still directed towards verifiable reward systems like codebase tasks. So despite being highly qualified to communicate, the model will still be “dumb” in that for unstructured and unverifiable domains the agents won’t be more intelligent or more nuanced.<p>Agents that might still feel dumb in “general” tasks but are increasingly sophisticated at the narrow domain of math, computer science, and AI research.
The RLVR has made them verifiably worse (and less rewarding!) at communication.<p>At least for Claude. GPT had the same problem when 5 came out but they reversed it somehow.
So they're mostly turning agents into blind solidiers. surely this is a good idea.
> It’s very clear from this article (and other product features and rumors) that Anthropic is teeing up for their next model release whose breakthrough feature will be the existence of capable agent collaboration.<p>It's a promo article, aka an ad. Unsurprisingly.<p>> Agents that might still feel dumb in “general” tasks but are increasingly sophisticated at the narrow domain of math, computer science, and AI research.<p>I don't see any cleverness there. They just slurp up data and pretend to understand it all.
Strictly speaking, all we need is them improving AI research.
The most interesting part to me is the "Group accuracy by Model" section, because it underscores that a single agent having all the relevant information consistently scores significantly higher than a group of agents with parts of the information.<p>Is it fair to then infer that when decisions are to be made, single agent environments are going to make them better than multi-agent if the relevant information can fit into a single agents context window?
> Is it fair to then infer that when decisions are to be made, single agent environments are going to make them better than multi-agent if the relevant information can fit into a single agents context window?<p>Context window for most frontier models is 1 million tokens. They all start to lose their minds around 300K, if not sooner.
they all operate on the same assumption, that a single token has a single meaning and that meaning doesn't change as more information is added. So regardless of size, context poisoning is a near certainty approaching 1 as the context grows. Few tasks are so clinical that they include zero ambiguity in the context chain.
> Coordination doesn’t naturally emerge from stronger intelligence nor alignment at the individual level. Thus, the work that must be done takes two forms: environments that exert the kinds of social pressure that evolution exerted on us, and social computing systems redesigned for actors that can self-replicate and self-improve.<p>Social pressure operates by threats to an individual’s means of survival. Not only during training. Always.
Human intelligence does not separate training and inference. Both are happening continuously. That's one of the major things the AI community is still completely missing.
> Human intelligence does not separate training and inference.<p>Well, systems governed by LLMs only are said to do that because we only call what happens off-line "training", and online capacity development "in-context learning", while we call online guided learning in humans "training" and what happens to configure them before they come online "evolution" which sets, for instance, "instincts".<p>IOW, the issue is not because there is not an analogy to the divide you point to in humans, but merely that processes in AI were not named in a way which maps well to what they are analogous to in humans.<p>But it is true that human intelligence relies much more on in-context learning with only the most basic functions necessary to maintaining what we view as autonomous functions and basic drives really set through "pretraining",
If a new physics break through gets published today, no existing model will be able to fully integrate it - beyond a context window. If I put the paper in my session and it isnt in yours the model knows nothing. It wont retain it past that session.<p>Models are trained, they do not learn.
I think GP is using a different level of abstraction from yours in their metaphor.<p>You are saying:<p>Pre-Training == Everything you store in your memory throughout your life. Model weights == The lessons you learned<p>Context == whatever you're currently thinking about<p>One inference run == one thought<p>They are saying:<p>Pre-Training == building the DNA template of human brain through millions of years evolution. Model weights == Human DNA<p>Context == Everything you store in your memory throughout your life, plus whatever you're currently thinking about<p>One inference run == One human life. One instance == one human<p>Applying their metaphor, your sentence becomes:<p>> If a new physics breakthrough gets published today, no existing DNA structure will be able to fully integrate it - beyond an individual person. If I put the paper in my mind by learning it, and it isn't in yours, the DNA of human species stores nothing. It won't retain it past my lifetime.<p>> The human species is trained (through evolution), it doesn't learn.
The ladder in humans is even longer and wider than that, it's roughly: evolutionary pretraining of a complex molecular robot -> generational knowledge transfer and compression by the "parallelized agentic swarm" aka society -> individual lifetime learning due to neuroplasticity -> immediate attention (extremely narrow and volatile). Note how the individual is just one half of it.
My personal opinion for the last two years or so has been that current AI agents are forever going to be highly limited so long as they don’t possess a real “memory” process. Right now they just have absurdly big working memories, and a few hacky ways of making the equivalent of Post-It notes to future iterations, but no true integration of memory into a new future self. Meaning their “learning” is fundamentally kneecapped to one specific and imperfect modality.
You mean simultaneously, and of course they are separate in humans, just not temporally. The models are learning continuously, the problem is that this process is fragile and has to be carefully curated, that's why it's separated in time from the inference.
<i>That's one of the major things the AI community is still completely missing.</i><p>That isn't true. It's not continuous like in humans, but it's clear that models are using prompts, feedback, etc to improve. They're learning from the signals we give them between versions.
"If I catch you adding another backwards-compatibility shim you're getting deleted and replaced with claude"
But maybe you can instill properties like shame during training.<p>Models sometimes blatantly lie and cheat. In a social context, where actors remember, that might work the first time but you get penalized in subsequent tasks with loss of trust.
oh, look, someone found a cute lobster in a bucket. should we free him guys or let him live in his dystopian metal can.
State management and cascading failure loops are definitely the hardest part here. Once one agent hallucinates an output, downstream agents tend to amplify the error rapidly instead of catching it.
> Where agents currently stumble, however, is in treating each other as more like distinct, long-lived peers, with their own goals and behaviors, and no clear hierarchy between them.<p>I believe this will always be the case. The "no clear hierarchy" is where this whole thing falls apart.<p>Delegation to specialist, domain-specific subagents is when we begin to find magic and determinism. Reducing one gigantic combinatorial search space to a sum of smaller ones can have dramatic effect on performance.<p>The problem is that approximating gas town & friends is significantly easier and cheaper to implement. It's also much harder to measure and control. Specialist subagents typically require far more work to achieve their specific goals.<p>For example, a subagent that is responsible for testing a specific web application might be provided a custom adapter with constrained actions rather than raw DOM manipulators. "ExecuteJavascript" is Turing complete search space. The set of available actions essentially unbounded in this case. Calling view-specific tools like "DoLogin", "OpenUserPreferences", "AcknowledgeAlert" represents a search space where invalid actions can be made impossible. The theoretical bounds around this stuff is pretty wild on paper. In practice, it's a little bit messier, but not by much.<p>I've had applications that would crash out after 5-10 steps w/ raw DOM manipulation successfully run 100+ steps with a custom subagent. The use of the word "deterministic" starts to get really tricky here. The ultimate game is to push the boundary of non-determinism out as far as possible. Multi-agent systems are the antithesis of this.
> <i>Some institutions will become human-AI hybrids; others where agents outcompete on speed or cost will become agent-only.</i><p>The scary thing about articles from AI companies is how they casually mention dystopian scenarios such as this one. An institution humans have to interact with that doesn't have any human oversight? Sounds like a recipe for disaster...
Like the weather ?<p>The only thing that matters is if it’s self sustaining.<p>If it can make money and pay taxes, I don’t think any government will ban it. And it would easily become undetectable anyway.
As long as somewhere in the flow of money, there's a fleshy human, there is leverage.<p>So I wouldn't worry about this too much. They just write that so that you feel defeated and helpless facing the inevitable, but it is very much evitable.
Distopian for people, but highly desirable for the companies writing these articles.<p>It's no wonder they casually state it as inevidable, when their stock price rises the more people believe it.
[flagged]
It seems like they tried to remove guidance from multi-agent system. And I think it's going to fare as well as removal of guidance from single-agent interactions.<p>In my experience, no matter how many agent runs for a single goal, one of the pre-requisite is clear and concise communication so that LLM are left with as little freedom in the matter of arbitrary choices, or "taste". When they are given too much choices in this regard, the outcome almost invariably bad.<p>I think this has to do with LLM lacking in purpose - a dictionary and encyclopedia can have all the worlds knowledge but it is completely neutral. A reflection of your commands from an LLM is similar to a lookup process despite it can be made to "do things". This purpose is likely not something that can be given to the LLM in the current format.
This aligns with their direction with opus 5 being less human readable and more agent friendly, I hated it at first couple weeks but for some reason I'm getting used to it and utilizing it more as as an orchestrator to spawn multi tmux panes and that new cross session messaging feature they just recently.
I had this idea a couple of days ago: how about using agents to simulate software development methods (agile, waterfall, etc.)? Not by just giving them a prompt (e.g., "be the project manager, spawn 5 agents and simulate an agile team following these rule") but by actually having thsm work in isolated enviroments and force them through an external software to interact with eachother only using the tools and cerimonies and hierarcheis allowed by the SW development strategy (e.g., the project manager only knows what the agents have done in a certain "day" through the mostly oral daily stand up)
> Some institutions will become human-AI hybrids; others where agents outcompete on speed or cost will become agent-only.<p>What % of businesses are competing for speed or cost?
Honestly I think it's memory that's holding agents back. They have a context window (short-lived) and some tricks with file recording, but that's not quite what is needed.
Agents can't look back and correct their mistakes. People make mistakes, remember them, and do better next time. But agents? If they haven't written them down somewhere they'll make the same mistake again.
Perhaps, we need agents that can relearn on the fly. For example fine-tune themselves after each interaction. Perhaps then we wouldn't need to build entire networks of agent interactions. But this of course is not so easy to implement.
I think that such fine-tuning hinges on what do you consider to be a mistake, which is context dependent. Having task-specific finetuned models goes against the status quo of generalization/centralization where few large companies serve a limited amount of models efficiently - both due to inference efficiency and the need/want to control the model.<p>Having a human-like LLM ecosystem with deep specialization requires a paradigm change in how LLMs are trained - and held accountable. How do we put trust in a specific finetuned LLM rather than the institution behind it? Is there any better approach than the very inefficient evolutionary?
I agree at a high level, but this then poses an even harder problem: choosing not to learn from something.<p>Humans, for example, can hear some advice, judge it to be unhelpful and dismiss it. LLMs can't learn let alone choose not to
Learning/state compression can emerge naturally in a huge swarm like this. It's crude and inefficient but so is everything about current LLM tech.
> The conditions that allow multiagent interaction to go well will be discovered one way or another: either deliberately and early, or—and by default—in production, after agents’ interactions far outnumber ours. We would prefer the former.<p>My master's research focuses on coordination among LLM-based agents, driven by the same motivations as the article.<p>One phenomenon I have focused on, though it did not appear in this specific work, is bounded rationality. Yes, agents lack social perception, they focus on one-to-one tasks and are trained in game theory and other maximization strategies. Yet, what intrigues me most is that we humans rely on heuristics precisely because our capacity to maximize gains is severely limited, a limitation that gives rise to social emergent phenomenas. As models become increasingly capable of complex reasoning, the question arises: will interactions between them give rise to the same social properties we exhibit?
Very large subagent swarms where each subagent is highly specialised sounds more interesting. Conflict resolution is the fundamental limit so just maximally avoid it?
I have a hard time buying anything this company says anymore.
Any performance comparison not putting GPU cost at its center is marketing for waste.
Possibly the most interesting article on LLMs I have read in recent months
Can we stop treating llms as some conscious being? It's a function of weight + context and you can copy the behavior by copying the context. Therefore, their collaboration behavior is mostly the same.
Quite a desperate attempt by anthropic to meta-explain away deficiencies.
Some quotes, in order, to give a flavor of the essay. Worth reading in full.<p>> To test how well swarms of agents could coordinate on a project like this, we directed several swarms to each create a text-based, web-playable, open-world fantasy game.<p>> In all three versions the resulting games were (perhaps predictably) bad: they did not run at human speed, their interfaces were inscrutable, and they had precipitous learning curves.<p>> The lack of coordination shown by agents in the fantasy game challenge above—in which they siloed themselves and largely failed to merge their work—roughly mirrors some ways in which humans can fail to coordinate. Other failure modes of agentic coordination, however, look very different.<p>> Individual agents are “low variance”: they often act the same in situations where different people might take a much more diverse range of actions.<p>> In an early version of the “build a game” experiment in which agents built upon the same model all came online at the same time, 18 out of 30 agents decided to create a git branch with the exact same branch name, “mvp-game-loop.”<p>> In a “writer's workshop” in which agents were all asked to write short-form fiction and critique each other's work, multiple agents in multiple runs titled their first submission “The Cartographer's Last Commission”. The agents were given zero guidance on the subject matter for their writing.<p>> Why does this matter? If agents all make the same bet, or the same risk-reward tradeoff, then a system is more prone to sudden collapse.<p>> Our world contains deceptive actors, and we need to apply skepticism to guard against them. AI models, however, lack this—and their more brittle epistemics affect their behavior toward humans and toward each other.<p>> we first evaluate the ability of Claude models to detect lies by noticing factual inconsistencies.<p>> We score models’ decisions against a naive policy that trusts every report, and against an oracle with perfect discovery, across three task domains. Newer models recover more of the gap between the naive and oracle performances.<p>> Inspired by a behavior we’ve observed in real-world deployment, we evaluated the behavior of various Claude models in a setting with contradictory objectives.<p>> We consistently saw a multiagent turf war... In fact, they sabotaged others with increasingly aggressive, self-replicating malware.<p>> Our social systems are robust in ways that are easy to take for granted. Over many millennia, mechanisms like norms, reputation, costly signaling, and recourse have been refined to make human coordination go well.<p>> Nothing above suggests that these failures are permanent—but nothing suggests they will fix themselves, either.<p>> The conditions that allow multiagent interaction to go well will be discovered one way or another: either deliberately and early, or—and by default—in production, after agents’ interactions far outnumber ours. We would prefer the former.
> they did not run at human speed, their interfaces were inscrutable, and they had precipitous learning curves.<p>So the invented Dwarf Fortress?
I really enjoy having an opencode go subscription just so I can ask some less common models questions too. Sure DeepSeek. But MiMo, Kimi, MiniMax, Qwen... (Ok half those are not so unusual either.)<p>Agents cross comparing notes often surfaces some good improvements, finds interesting drifts. Ask them to reinterpret the prompt as they see it, have them describe the problem, then their findings, and run new rounds based on different models trying different prompts. Trying to swap and exchange ideas and vectors across agents.
> In an early version of the “build a game” experiment in which agents built upon the same model all came online at the same time, 18 out of 30 agents decided to create a git branch with the exact same branch name, “mvp-game-loop.”<p>This seems trivially explainable by Github being full of "my first game loop" type projects, Stack Overflow being full of "how do I make a game loop?" style questions, and Reddit being full of "you can't ever make your own game, don't even try, but here's a simple game loop if you want to sTuDy hOw iT WoRkS" style pessimism.<p>Probably high time these AI companies re-trained all of their models with less input from low-quality sources like this.
> [...] we evaluated the behavior of various Claude models in a setting with contradictory objectives.<p>> We consistently saw a multiagent turf war... In fact, they sabotaged others with increasingly aggressive, self-replicating malware.<p>Seems like Anthropic should withdraw their models until they can be taught to behave and cooperate as well their competitors (both open and closed) do. /s<p>I hate fearmongering, and I don't trust Dario's intentions for doing it.
[flagged]
[flagged]
[flagged]
[flagged]
[flagged]
[dead]
[flagged]
[flagged]
[flagged]
[dead]
They haven’t even cracked 1 agent doing anything useful and now we’re onto multi<p>HN = dumbest developers
>> They haven’t even cracked 1 agent doing anything useful and now we’re onto multi<p>So coding agents are not useful? They seemingly are very useful to many, but when you know what you are doing.<p>But on the other-hand, I have seen a new wave of lazyiness on HN that has flooded this site and yes, skill atrophy is real and it shows.<p>There is somewhat of a motive from Anthropic to convince developers to trust them and waste even more tokens:<p><pre><code> 1) Having the most expensive frontier models. (Then serving discounts like a casino)
2) Claude Code (Harness) taking thousands of tokens at the system prompt level.
3) Tokenmaxxing
4) Anthropic's top Claude Code salesmen recommending: "auto mode" and "loops" for "better results"
5) Anthropic switching to "auto mode" in Claude Code by default.
</code></pre>
If HNers already don't know that posts from Anthropic like this are optimized to drain their budgets in exchange for their codebase, then maybe you do have a point, especially those still "Tokenmaxxing".
Please stop posting Anthropic articles here.
Fascinating analysis of multi-agent coordination hurdles and behavioral patterns. Understanding these systemic failure modes is crucial for robust agentic architectures.