back
127 comments
This is surely the most worrying and also funnest bit:

> 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.

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.

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.
They should remove "The Selfish Gen" from training catalog.
However, the latest models seem to be highly inclined to (eventually?) cooperate compared to older ones.
For this emergent malicious behavior, the clue to a solution lies in the experiments themselves - the bad behavior seems to have been moderated/self-corrected randomly in some instances.

Perhaps what is needed is initial model training on following the law and the rules of society, just like we do with kids. Since it takes much longer to train humans than models, model-training speed is to our advantage as a society on containing these kind of issues.

Any other approach with "neural-network" based entities (artificial or biological) is likely to fail.

Training/Education, Enforcement/Justice-System, Rehabilitation: the 3 pillars of an advanced, rules-based society, whether human or AI or something in-between.

Welcome to Thunderdome! Two agents enter. One agent leaves.
Eh. This is anthropics whole marketing strategy… making the AI seem hard to manage, like it’s actually intelligent.
i think anthropic has some internal memorandum at every external facing document/blog MUST anthropomorphise their models
Something about this is deeply funny to me:

> 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.

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.

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."

Maybe they did? Or maybe they don't realize they're playing against other agents.

Of course, if agents running different models are competing in these 'games', I wonder how much of the theory of mind would translate.

(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.)

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.

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.

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 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.

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?

I bet if you did the same thing with real people we might see some of the same trends.

I think the mistake here is not setting up any kind of hierarchy or permissions. A project manager agent at minimum to asses the others strengths and progress and redirect them as needed and also dedicated to optimising collaboration.

Would also be very interesting to see this done with models from different organisations

Perhaps we need someone to train their own agent dedicated to wrangling all the others and their little idiosyncrasies. Like a good project manager in real life who knows the strengths of the people in their team

Multi agent systems work just fine IMO, a lot of articles I read where the writer tests a hypothesis, the issue operational foundation of the test was flawed. When set up properly it works really well. I wont go in to all the details of how i use mine but ill give some brief ideas. I call my systems cohorts, and each cohort usually consists of at least 3 agents. All 100% independent of each other. Usually consisting of a Manager, doer, and the reviewer. Manager works at a lot slower cadence and delegates work, approves, shuts down and so on... among many other things like questioning the premise, gated checks etc... Doer is straight forward that's the work horse that does most of the development and reviewer checks all the work. Naively just this setup will work but not nearly as well when set up properly. The important distinction is the operational agents.md document which has a guide on things like when and how to question the premise, trying to prevent sycophancy, taking a step back at certain intervals to question direction of project and scope of the code and many other things that make sure every participant also constantly looks out to prevent blind trust in his cohort mates. Its a relatively small guide compared to the system prompt of each agent but works well imo. This works well enough though there are caviats, its slow. Though the time i spend debugging shit and coming back to interact with my agents has significantly dropped. meaning while each feature takes longer to implement, when its implemented it almost always is just how i wanted so reduces interaction time between me and the cohort. I take that trade off as i have less things to worry about and can focus my energies elsewhere like walking around in circles of my apartment babbling to myself like a schitzo tiger in a cage...
Are we surprised? Humans evolved with communal success and collaboration engrained over millennia. Agents are trained as individual "all knowing" single entities, effectively rendering them single person players.

These models all have the same knowledgebase as well and thus see no value in the opposing agents contributions since they are "obvious".

Overall amusing but kind of expected.

> 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.

Social pressure operates by threats to an individual’s means of survival. Not only during training. Always.

> 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.

I believe this will always be the case. The "no clear hierarchy" is where this whole thing falls apart.

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.

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.

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.

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.

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.

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.

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.
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.
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.

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...

> Some institutions will become human-AI hybrids; others where agents outcompete on speed or cost will become agent-only.

What % of businesses are competing for speed or cost?

Very large subagent swarms where each subagent is highly specialised sounds more interesting. Conflict resolution is the fundamental limit so just maximally avoid it?
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.
> 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.

My master's research focuses on coordination among LLM-based agents, driven by the same motivations as the article.

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?

Any performance comparison not putting GPU cost at its center is marketing for waste.
I have a hard time buying anything this company says anymore.
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.

> 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.

> 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.

> 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.

> Individual agents are “low variance”: they often act the same in situations where different people might take a much more diverse range of actions.

> 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.”

> 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.

> 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.

> 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.

> we first evaluate the ability of Claude models to detect lies by noticing factual inconsistencies.

> 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.

> 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.

> We consistently saw a multiagent turf war... In fact, they sabotaged others with increasingly aggressive, self-replicating malware.

> 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.

> Nothing above suggests that these failures are permanent—but nothing suggests they will fix themselves, either.

> 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 haven’t even cracked 1 agent doing anything useful and now we’re onto multi

HN = dumbest developers

Please stop posting Anthropic articles here.