What it was for
Nothing yet, which is the point of the verdict. The agent team is being proposed rather than inherited, and it is being proposed in the shape of the thing it would replace. Zhenfeng Cao's 2026 paper argues that AI agents restructure software work outright, and its Stage III (2026–2029) is "multi-agent teams": coordinated swarms, role specialisation, shared memory, humans as PM, architect and auditor. Agents as digital team members, each with a role and a handoff.
That shape is not neutral. It is the org chart, and the org chart is an answer to a constraint agents do not have — a person can only hold so much, so you split the problem and pay a coordination cost to put it back together.
The verdict
EMERGES — and the entry exists to say which shape emerges, because the evidence separates two things the word "multi-agent" bundles together.
Handoff. Four agents in a relay, each compressing what it knows into a paragraph for the next. Measured live in the PSG demo: roughly five times the tokens and five times the wait, for an answer indistinguishable from the single agent's. Four opinions, or one opinion billed four times.
Shared state. The same four agents reading and writing the same typed facts, with no paragraph in between, so there is no boundary to compress across in the first place.
The distinction is not stylistic. It decides whether adding an agent adds an opinion or adds an interest payment.
What changes
The strongest evidence is not on either side of that split. It is about what happens to any agent over a long enough sequence.
EvoClaw (Deng et al., March 2026) evaluates agents on dependent sequences of milestone tasks drawn from real repository histories, rather than on isolated fixes. Frontier agents score above 80% on isolated tasks and at most 38.03% across the sequences, over seven repositories and four harnesses including Claude Code and Codex. The named failure modes are error propagation and regression: each step inherits what the last one got wrong.
That is the relay result again at a different scale. The cost is not in the model. It is at every boundary where something has to be compressed, handed on, and trusted.
Read the other side honestly. The most-quoted number for orchestration, a 93% cut in root-cause identification time, comes from a Cisco internal pilot written up on LangChain's blog: more than twenty debugging workflows over one month, self-reported, which is a vendor write-up rather than research. It is also the shared-state shape rather than the relay, with parallel investigation, shared context and cross-validated findings. If it holds, it corroborates the distinction rather than the roadmap built on top of it. The same caution applies to the paper's other figures, which characterise 2026 capability with a SWE-bench Verified score from November 2024.
The strongest objection
That this is a temporary limit. Cao says so directly: the gap "is not fundamental — it reflects limitations in context management, memory architecture, and verification mechanisms." Better memory, better context, and the relay stops leaking.
Partly conceded. Context windows have grown before and will grow again. But the objection concedes the claim it means to answer: if the fix is memory architecture, then the unit of design is state rather than headcount, which is what the verdict says. More agents will not repair a boundary that is still there to be crossed.
Falsification: downgrade to TRANSFORMS if a relay topology beats a shared-state topology of the same agent count on a dependent task sequence, on a published benchmark rather than a vendor pilot.