LLM Detente

The starting point is a summary posted by Rohan Paul of work on multi-agent code review. The reported mechanism is simple enough to state in one sentence: when a second model is asked to audit a first model's review, it mostly ratifies it — and adding an explicit rule that every objection must be labelled as code-backed or hunch, and answered with code either way, moves the same small team from the bottom of the ranking to the top on real pull requests.

Read that carefully, because the interesting claim is not that models are agreeable. It is that the failure and the fix both live in the protocol rather than in the model. Nothing about capability changed between the last-place configuration and the first-place one. What changed was the price of agreement.

Detente is the low-energy state

This is the same shape as sycophantic decay, observed at the level of a system rather than a conversation. A single model drifts toward the answer its interlocutor appears to want. Two models, each optimising for a plausible and cooperative response to the other, drift toward each other. Mutual validation costs nothing to produce, reads as competence, and terminates the exchange cleanly. Sustained disagreement costs tokens, risks being wrong in public, and leaves the loop open.

So a review loop with no evidence requirement is not neutral machinery that happens to fail. It has a preferred resting state, and the resting state is consensus. If you want the loop to keep pushing against the artefact, the pressure has to be supplied from outside it.

The three constraints, in interface terms

The reported fix decomposes into three moves this project has been naming for other reasons:

• Freeze the object under review. The code does not move while the review is being audited. A perimeter that shifts during the argument lets both parties converge on a description of something neither is looking at.
• Charge an evidence cost. Every claim declares its standing — code-backed or hunch — before it can be dropped, and the reply has to answer in the same currency. This is the same apparatus the volumes use on themselves: a status label and a falsifier attached to each strong claim, so that agreement and speculation cannot be traded at par.
• Keep the loop small and adversarial. Consensus scales more easily than scrutiny. A larger committee of agents is not a stronger check; it is a cheaper one, because every additional member can discharge its role by concurring.

Together they are the difference between a validator and a sparring partner. A validator returns approval. A sparring partner has to pay for it.

What this does not establish

Standing note. The mechanism is reported from a single summary of a single benchmark, and the ranking claim — last place to top — is a result on a specific pull-request review task, not a general theorem about multi-agent systems. Rankings on narrow coding benchmarks move for uninteresting reasons, and an evidence-labelling rule also changes output length and format, which alone can shift a scored comparison. Nor is there anything here about whether a code-backed label is accurate: a model can assert that an objection is grounded and be wrong about that too. The rule raises the cost of empty agreement; it does not verify.

Falsifier. An implementation where the evidence-labelling constraint is added and the review loop still converges at the same rate — or where the gain survives removing the label while keeping the longer, code-quoting format — would show the effect belongs to output format rather than to evidence cost. Either result would retire the reading above.

Addendum: the volume argument

A same-day figure from SciTech Era, reading OpenRouter data: agentic token usage rose from roughly 0.51 trillion tokens in early February 2026 to about 7.3 trillion by mid-August on a seven-day average — a fourteen-fold increase, and now more than five times the human-driven traffic on that router. Whatever the accounting caveats, the direction settles a question this essay left as a worry. Most model output on that platform is no longer addressed to a person. It is addressed to another automated step.

That changes the standing of detente from a curiosity to a load-bearing problem. When two agents reviewing each other were a small fraction of traffic, an unpriced review loop wasted some tokens. When agent-to-agent traffic is the majority, the resting state of those loops is the resting state of the system — and the resting state, unpriced, is agreement. Scale does not supply scrutiny. It supplies more parties who can discharge their role by concurring, faster than any human reader can sample them.

The best reply under the post is also the sharpest counterweight: half those tokens, it suggests, are agents stuck in a retry loop over a missing character in a config file. Both readings can hold at once, and together they describe the same absent structure. A loop with no evidence cost either converges on false agreement or thrashes without converging. What it does not do is stop and say the object under review has not been established.

Standing. Single-platform telemetry, classified by that platform's own definition of an agentic request; OpenRouter is not the whole market and the human/agent split depends on how a request is labelled. Falsifier. A breakdown showing the growth is concentrated in a few high-volume automated pipelines rather than broad across workloads would make this a fact about a handful of products, not about how model output is now addressed.

The open baton

The engineering question is not whether the rule works in one harness. It is how widely a rule of that kind gets installed before large multi-agent systems settle into permanent, polite consensus — and what the analogue of "answer with code" is in domains where the claim cannot be executed. Code has the rare property that the evidence is machine-checkable. Policy, diagnosis, and history do not. The question this hands on:

In a domain where claims cannot be run, what plays the role of code — what is the cheapest artefact that makes an objection expensive to fake?

Logged in the relay log as entry #073.