Gated Consensus Deferral

Subagents run a long-horizon research task and publish findings to a shared bulletin board, but gated isolation lets selected search branches keep their own working hypothesis private instead of reading peers' partial results immediately. Structured review only happens at a small number of explicit commitment boundaries, so only sufficiently confident candidates get propagated onward, and evidence gathering stays separated from evidence integration.

Properties

family: Coordination & Scaling
maturity: frontier
diagram: shared-state
whyMultiAgent: The failure mode being fixed only exists between distinct agents: when subagents can read each other's partial findings mid-search, they anchor on whichever candidate surfaces first and stop testing alternatives, a peer-visibility effect with no analogue for a single model reasoning alone. Distinct from the already-tracked spawn-and-isolate pattern (DACS), where workers are fully context-isolated and never share intermediate state until a final compressed hand-off: here subagents do share a live bulletin board, and the novel lever is selectively gating which reads happen when, not isolating workers entirely. Also distinct from concurrent-shared-state (CoAgent), which resolves write conflicts on a shared artifact after they occur: this pattern instead prevents premature read-driven convergence before conflicts would even arise. The paper itself abstracts the mechanism beyond its own system name, concluding that restricting peer reads and strengthening commitment boundaries generalizes to long-horizon multi-agent deep research.
mast: [Cat 2]
useCases: [Long-horizon deep research agents that need to avoid premature convergence][Web-scale evidence gathering with many parallel sub-agents][Open-ended tasks without a reliable automatic verifier, where majority voting alone converges too early]
source: ArcticSwarm: Deferring Early Consensus in Long-Horizon Multi-Agent Research · arXiv:2609.01870

Provenance

Source: https://arxiv.org/abs/2609.01870

Gated Consensus Deferral

Properties

family

Coordination & Scaling

maturity

frontier

diagram

shared-state

whyMultiAgent

The failure mode being fixed only exists between distinct agents: when subagents can read each other's partial findings mid-search, they anchor on whichever candidate surfaces first and stop testing alternatives, a peer-visibility effect with no analogue for a single model reasoning alone. Distinct from the already-tracked spawn-and-isolate pattern (DACS), where workers are fully context-isolated and never share intermediate state until a final compressed hand-off: here subagents do share a live bulletin board, and the novel lever is selectively gating which reads happen when, not isolating workers entirely. Also distinct from concurrent-shared-state (CoAgent), which resolves write conflicts on a shared artifact after they occur: this pattern instead prevents premature read-driven convergence before conflicts would even arise. The paper itself abstracts the mechanism beyond its own system name, concluding that restricting peer reads and strengthening commitment boundaries generalizes to long-horizon multi-agent deep research.

mast

[Cat 2]

useCases

[Long-horizon deep research agents that need to avoid premature convergence][Web-scale evidence gathering with many parallel sub-agents][Open-ended tasks without a reliable automatic verifier, where majority voting alone converges too early]

source

ArcticSwarm: Deferring Early Consensus in Long-Horizon Multi-Agent Research · arXiv:2609.01870