More agents do not automatically produce independent judgment

Anthropic’s February 2026 compiler experiment hit a bottleneck when agents encountered the same Linux build failure and interfered with one another’s fixes. Comparing smaller portions against GCC helped divide the problem. The agents needed an external reference, not simply more discussion. Anthropic C compiler experiment.

An agent that challenges another agent can produce a persuasive exchange. That exchange may help expose assumptions, but its value depends on what actually differs between the participants. If both rely on the same incomplete evidence, a debate can reach confident agreement without discovering the missing fact. Independence must be designed rather than implied by role names.

CAMEL, first submitted March 31, 2023, explored role-playing and cooperation between communicative agents. It provided a framework for studying such interactions. The paper's research setting does not establish that a simulated critic has the independent knowledge or authority of a qualified external reviewer. CAMEL paper.

Give the critic a different job

Our interpretation is that a reviewer should have a distinct test to perform. It might verify citations against original documents, run a separate test case, or inspect an authorization boundary. Asking it to reconsider the same prose is a weaker intervention. The review becomes more informative when it can return evidence the author did not already use.

AutoGen's August 16, 2023, paper described agents that could combine language models, tools, and human input. That flexibility matters because diversity can come from the resources and checks assigned to a role, not only from the wording of its system instructions. AutoGen paper.

Preserve disagreement

We would retain unresolved disagreements in the final handoff and identify what evidence would settle them. A manager agent should not collapse uncertainty into a clean answer merely to finish the conversation. Multi-agent review earns its cost when it catches a specific error or strengthens a specific claim. The number of voices is a poor substitute for that observable result.

Why agreement is weak evidence

The compiler account is a useful counterweight to agent-team demonstrations built around multiple conversational roles. Agreement between several model instances can look like verification, but the instances may share the same missing fact, misleading instruction, or flawed interpretation. If every participant reads the same summary, repetition can increase confidence without adding information. The number of approving messages is therefore a poor measure of how thoroughly a claim was checked.

This matters in software because a reviewer can repeat the author’s explanation without exercising the changed path. It matters in research because a second agent can cite the first agent’s summary rather than inspect the underlying document. In either case, the workflow has added a voice while leaving the original evidence unchanged. Useful independence comes from separate observations or a different test, not from a different role name.

Anthropic’s approval design separates the evidence

In its March 25, 2026 description of Claude Code auto mode, Anthropic said the action classifier sees user messages and tool calls while excluding the acting model’s own prose and tool outputs. The stated aim was to reduce the ability of the agent’s explanations—or hostile material it had read—to persuade the checker. This is a specific example of separating a reviewer’s evidence from the author’s narrative, rather than asking another model to endorse the same reasoning. Anthropic Claude Code auto mode.

Debate and verification answer different questions

The older CAMEL and AutoGen work discussed below explores how role-based interaction can organize tasks. That can be useful for generating alternatives, exposing assumptions, and dividing effort. It does not imply that a consensus has the statistical meaning of independent human measurements. Models with similar training and context can make correlated errors, and even different models can be led astray by the same false premise.

The more productive use of disagreement is to turn it into a testable question. If two agents disagree about whether a function handles empty input, execute the case. If they disagree about a public announcement, inspect the original language and date. If neither can resolve the question, preserve the uncertainty rather than manufacturing a majority verdict. This keeps debate connected to the environment instead of letting the conversation become its own evidence.

The external check is the important third party

In the compiler story, GCC supplied a reference outside the agents’ shared conversation. That does not make every comparison effortless or every test complete. It changes the source of authority: a concrete behavior can contradict a persuasive explanation. The same principle applies when an agent checks a database record, renders a document, or observes a service after deployment.

The implication for agent products is subtle. A team interface can be valuable, but its strongest feature may be the diversity of checks it coordinates rather than the number of agents it displays. One agent reading the specification, another exercising the implementation, and a third comparing the two can supply meaningfully different evidence. Three agents reading the same generated description may not.

The public reports support a more restrained view of collective intelligence. Multiple agents can increase search breadth and parallel effort. They do not abolish the need for an independent source of truth. The systems most worth watching will make that source visible and explain how disagreement changes the work. Their credibility will come from what they can disconfirm, not merely from how smoothly their participants agree.

The research behind the release

CAMEL’s March 2023 research used role-playing and inception prompting to study cooperation between conversational agents. Those roles can structure an exchange and encourage participants to pursue complementary tasks. They do not create separate sensory access to the world by themselves. In a deployed system, a reviewer becomes more informative when it reads the original requirement or executes a test that the author did not use. The important difference is the evidence available to each participant. That is why a small team with distinct checks can be more valuable than a large discussion organized around identical information. Original research paper.