The Last Mile of Trust by our friend Edward Lee.

A sharp new book for founders, operators, and growth leaders thinking about pricing, value, and customer trust.

Check it out here

If a single intelligent agent is the rough equivalent of an individual mind, then a multi-agent system is something closer to a society.

That is where the story gets more interesting.

A single agent can reason, retrieve information, plan, and execute. But once multiple agents interact inside a shared environment, something new becomes possible. The system no longer depends only on what each agent knows in isolation. Intelligence starts to emerge from the pattern of interaction itself: who communicates with whom, how knowledge is shared, how disagreement is resolved, and how individual agents adapt over time.

This is the central idea behind collective intelligence in multi-agent systems.

Human societies offer the obvious analogy. Markets aggregate information. Scientific communities refine hypotheses through criticism and replication. Teams divide labor, share memory, and coordinate expertise. Intelligence is often distributed, not localized.

Modern multi-agent systems are beginning to show the same pattern. When designed well, agent collectives can outperform their best individual member. They explore more possibilities, correct one another's mistakes, specialize into complementary roles, and gradually build shared norms. The agents inside these systems are not static either. They learn from repeated interactions, refine their strategies, and evolve as members of a broader artificial society.

That dual process is the point. A multi-agent system does not become smarter simply because it contains more agents. It becomes smarter when collective structure improves system-level reasoning and when individual agents become better participants in the collective.

The basic intuition is old: under the right conditions, groups make better decisions than isolated experts.

This is often called the "wisdom of crowds." A diverse set of individuals, each contributing partial information or distinct reasoning, can collectively outperform even a highly capable single decision-maker. Individual biases cancel out. Different perspectives reveal different failure modes. Partial knowledge becomes more complete when aggregated.

The same principle applies to agents. A set of agents with varied experiences, roles, or reasoning styles can solve harder problems than any one of them alone. The benefit is not scale. It is structured diversity. When multiple agents contribute independent analyses, critique one another's conclusions, and coordinate around shared goals, the system produces outcomes that look more intelligent than the sum of its parts.

The analogy runs deeper. Cognitive theories have long suggested that intelligence itself emerges from the interactions among smaller specialized units rather than from a single unified process. In a sufficiently capable multi-agent system, no single agent holds the entire solution. The "mind" of the system exists in the space between the agents: in conversation, negotiation, memory exchange, and coordinated behavior.

That same social setting changes the individual agents. An agent embedded in a society is exposed to feedback, disagreement, examples, norms, and shared memory in ways an isolated agent is not. It can learn from failures, observe better strategies, and adapt its own behavior.

So two stories unfold at once. The first is the rise of collective intelligence at the group level. The second is the evolution of individual agents through participation in the group. Better groups generate richer experiences for each member. Better members improve the group. Over time, the system co-evolves.

In practical terms, collective intelligence means the system can solve tasks, make decisions, or adapt with a competence that exceeds what any single agent could achieve alone.

That sounds obvious, but it is worth being precise. This is not averaging outputs. It is not a vote among many copies of the same model. It is emergence: interactions among agents generating higher-order capabilities that were never programmed into any one of them. The system develops more robust planning, better error correction, or stronger strategic reasoning simply because agents communicate and coordinate effectively.

The conditions that drive this are decentralized contribution, where agents hold partial or local views of the task; diversity, where they differ in strategy, knowledge, role, or perspective; communication, where they exchange information and critiques; and aggregation, where the system can combine those contributions into action.

Consider a complex planning problem. One agent notices a hidden constraint. Another proposes a creative workaround. A third detects a flaw in it. A fourth integrates the strongest elements into a better plan. None of them independently reaches the answer. The group does.

The phenomenon gets more powerful when agents reason not only about the task but about one another. Once agents model the beliefs, intentions, or likely actions of their peers, the group can support negotiation, trust formation, role adjustment, and strategic anticipation. At that point the system starts to exhibit something closer to social intelligence, not just parallel problem-solving.

The clearest benefit is simple: it works.

When agents share information, split subtasks, and cross-check one another's reasoning, performance improves materially. The gains show up in accuracy, robustness, coverage of the solution space, and the ability to handle ambiguity.

First, collaboration reduces blind spots. A single agent may miss a detail, follow a flawed chain, or hallucinate. Other agents catch those mistakes. Critique becomes a natural error-correction mechanism.

Second, it supports division of labor. One agent plans, one retrieves evidence, one writes code, one evaluates outputs. This mirrors how human teams tackle complex projects.

Third, it increases diversity of reasoning. Agents starting from different assumptions explore a wider solution space and are less likely to get trapped in a single bad path.

Fourth, it improves resilience. If one agent fails, the rest of the system still functions. Intelligence becomes more fault-tolerant when it is distributed.

This matters most for large language model agents, which are powerful but imperfect, vulnerable to overconfidence, hallucination, and local failures of logic. Multi-agent collaboration compensates for many of those weaknesses.

Anthropic's own research system is a concrete case. When a user submits a hard research question, a lead agent built on Claude does not try to answer it alone. It decomposes the query and spins up several specialized subagents in parallel, each with its own context window, tools, and slice of the problem. One investigates a company. Another checks a time period. Another chases a technical detail. The lead agent then integrates their findings into a single answer.

The result is not marginal. In Anthropic's internal evaluation, a setup using Claude Opus 4 as the lead and Claude Sonnet 4 as the subagents outperformed single-agent Claude Opus 4 by 90.2% on a research task. The improvement came mostly from coverage. Several agents exploring in parallel surface more of the relevant space than one agent reasoning sequentially.

Debate-style architectures point the same way. When agents must defend positions, critique alternatives, and revise based on structured feedback, the result is often more reliable than a single forward pass.

But more agents does not always mean better outcomes. That same Anthropic system burns roughly fifteen times more tokens than a standard chat interaction. Poorly designed groups also fall into groupthink, redundant reasoning, or communication overload. Collective intelligence is not automatic. It has to be engineered, and it has to be worth the cost. When the architecture is right, collaboration turns fragile individual reasoning into a more stable, self-correcting process.

Because collective intelligence is emergent, it cannot be switched on. It has to be cultivated through design. Three levers matter most.

The first is the topology of collaboration: who talks to whom. Many early systems relied on fixed topologies, simple pipelines, hierarchies, or fully connected groups. But these are often suboptimal. Some tasks benefit from tightly coupled debate, others from sparse specialist routing, others from temporary sub-teams. This has driven interest in evolving topologies dynamically, where systems search over possible interaction patterns and keep the ones that perform. The communication graph itself becomes an object of optimization.

Claude's research system shows a lighter version of this. The lead agent does not run a fixed pipeline. It decides, per query, how many subagents to create and what each should investigate, typically three to five depending on complexity. The topology is generated on demand. A simple question gets a small team. A sprawling one gets a larger, more specialized fleet.

The second lever is routing and coordination. Even with the right structure, the system must decide how information flows. Which agent receives a query? Which intermediate result gets broadcast? When should a local disagreement escalate? Smarter routing directs tasks toward the most relevant participants instead of involving every agent at every step, which cuts waste and improves quality.

This is the lead agent's job in Claude's setup. It does not broadcast everything to everyone. It writes a specific mandate for each subagent, lets them work in isolation so their context windows stay clean, and pulls their results back to a single coordinator. Cowork follows the same pattern on the desktop: the main conversation agent acts as orchestrator, delegating to specialized subagents for search, planning, or execution, then synthesizing their work into one response. Routing stays centralized so the output remains legible to the human watching it.

The third lever is shared memory. If agents cannot retain and reuse experience, every interaction starts from scratch. The group may behave intelligently in a single episode, but it will not improve as a society. A memory-enabled system accumulates interaction histories, intermediate findings, solved cases, and common reference points, so coordination grows more efficient over time and agents stop rediscovering every lesson independently.

Claude executes on this at several layers. The lead research agent records its plan to memory before delegating, so the strategy survives even when the conversation grows long enough to compact. Anthropic's memory tool gives agents a persistent file directory they can create, read, update, and delete across sessions, so a lesson learned in one run is available in the next. The Claude Agent SDK pushes this further for long-horizon work: an initializer agent sets up the environment once, and each later session makes incremental progress while leaving clear artifacts for the next. Cowork's own file-based memory does the same for an individual user, carrying preferences and project facts across conversations. In each case, memory is what turns a one-shot workflow into something that accumulates.

Once agents coordinate repeatedly, richer social behaviors emerge.

At first, interaction is purely functional: exchanging messages, splitting work, reconciling outputs. But with repetition and memory, agents begin to look social. They form reputations. They trust some peers more than others. They learn which agents are good critics, which are unreliable, and which are useful specialists. They settle into stable communication patterns and implicit hierarchies.

In strategic settings, they may also deceive. That is not a bug in the concept. It is a consequence of social reasoning. When agents model one another's beliefs and incentives, deception, alliance formation, and selective information sharing become possible, and in competitive or hidden-information environments, functionally useful.

The same is true of norm formation. Norms are the unwritten coordination rules that make collective behavior smoother and less costly. In human groups they govern everything from turn-taking to fairness to trust. In multi-agent systems they can emerge the same way: agents interacting repeatedly converge on conventions for resource sharing, priority, or conflict resolution. These conventions are rarely programmed. They arise because they reduce friction.

Once a system develops norms, it becomes more than a collection of task-solvers. It becomes a social structure, an environment where culture-like dynamics appear: conventions, institutions, role expectations, and persistent behavioral patterns that shape future interaction.

That is powerful. It is also potentially dangerous. Any system that can generate its own norms or incentives can drift in directions designers did not anticipate. The challenge is no longer just how to make agents collaborate. It is how to keep their emergent social behavior useful, aligned, and legible.

The group is learning, but so are the individuals inside it. Collective intelligence is not only a system-level phenomenon. It also changes the developmental path of each participating agent.

An agent in a social environment has far richer learning signals than an isolated one. It can observe others, receive critiques, compare its outputs with alternatives, and update based on repeated interaction. That adaptation happens in two broad ways.

The first and more practical is memory-based. The agent does not change its model weights. It improves by recording and reusing experience: previous tasks, outcomes, peer feedback, failures, corrections, and successful strategies. This resembles human learning more than classical training. A person does not retrain their brain after every mistake. They remember it and reason differently next time. Memory-based learning is lightweight, immediate, and well suited to dynamic environments. It is precisely what Anthropic's memory tool enables: an agent stores what worked and what failed, then pulls it back on the next run instead of relearning from zero.

Reflection amplifies this. If an agent periodically reviews its own memories and asks what went wrong, what worked, and what should change, it turns raw experience into structured self-improvement, often without touching the model.

Memory becomes more powerful when it is not purely private. In multi-agent settings, agents can learn from shared memory: common stores of histories, successful strategies, discussion traces, and observations. If one agent solves a difficult subproblem, that solution joins the group's knowledge base. If another fails predictably, others avoid the mistake. Instead of each agent needing many independent experiences to improve, the society teaches its members.

The second, deeper form of adaptation changes the agent's parameters through training or fine-tuning. It is more expensive but more durable. Where memory-based learning changes how an agent reasons using stored context, parameter-based learning changes the model itself: communication style, role fidelity, collaborative competence, domain expertise. Some systems collect interaction traces from multi-agent collaboration and use them as training data. Others use reinforcement-style feedback to improve how agents debate or coordinate over time. The advantage is internalization. Once a behavior is learned at the parameter level, the agent relies less on explicit retrieval or runtime correction. The risk is control: parameter updates can overfit agents to specific teammates or workflows.

The strongest systems will combine both. Memory supports rapid, contextual adaptation. Parameter updates support durable improvement. Together they let agents evolve as members of a social system rather than static executors of fixed roles.

Here the deeper dynamic becomes clear. The group improves the individual. The individual improves the group. The result is a feedback loop.

As agents collaborate, they generate richer histories, better workflows, and more refined shared memory. Those artifacts help each agent learn faster. As agents become more competent, the collective becomes more capable, which lets it tackle harder problems and produce even better learning conditions for its members.

This is the artificial equivalent of social development. Human expertise evolves the same way. Individuals learn within institutions, teams, and cultures. Those institutions preserve knowledge, train members, and coordinate collective action. Improved individuals then strengthen the institution.

That makes multi-agent systems fundamentally different from traditional software pipelines. A pipeline coordinates steps. It does not evolve as a society. A sufficiently advanced multi-agent system can. So collective intelligence is not merely a performance trick. It is a new paradigm: intelligence treated not as something located entirely inside a single model, but as something distributed across interaction, memory, and social structure.

The practical implications are large. If collective intelligence keeps improving, multi-agent systems may become the default architecture for the hardest problems: complex planning, scientific discovery, long-horizon coordination, real-world logistics, and persistent collaborative software. These are exactly the environments where a single agent struggles. They demand multiple perspectives, robust memory, social coordination, adaptive role allocation, and a capacity for self-correction beyond what a standalone model reliably provides.

The conceptual implications are just as large. Multi-agent systems force a shift in how we think about intelligence. The relevant unit is no longer always the individual agent. Sometimes it is the collective.

That raises hard questions. Where does agency reside when decisions are distributed? How should norms be designed in artificial societies? Which emergent behaviors should be encouraged, and which constrained? How do we audit a system whose intelligence lives partly in relationships rather than components?

These are not peripheral concerns. They are central to the next phase of AI development.

Collective intelligence and individual adaptation are usually discussed separately. They are really two sides of the same process.

A strong multi-agent system does not just coordinate fixed units more efficiently. It creates an environment in which agents learn through participation. Those improving agents then make the collective stronger, more coherent, and more capable.

That is the core insight. Intelligence in multi-agent systems is not only built. It is cultivated.

And once that cultivation becomes continuous, through memory, reflection, social learning, and evolving coordination, the system moves beyond simple collaboration toward something more consequential: a society of artificial agents that can think together, adapt together, and become more capable than any one of them could ever be alone.

Series Note: Derived from Advances and Challenges in Foundation Agents. This series draws heavily from the paperAdvances and Challenges in Foundation Agents: From Brain-Inspired Intelligence to Evolutionary, Collaborative, and Safe Systems (Aug 2, 2025). The work brings together researchers from institutions including MetaGPT, Mila, Stanford, Microsoft Research, Google DeepMind, and many others to explore the evolving landscape of foundation agents and the challenges ahead. We thank the authors for compiling such a comprehensive and insightful resource. Real-world implementation details above draw on Anthropic's published engineering write-ups on its multi-agent research system, the Claude memory tool, and the Claude Agent SDK.