Replace the orchestrator with a shared queue and the swarm gets faster and more accurate
What is the main agent costing you while it waits?
Opening
Parallel subagents promise throughput, but someone still coordinates them, and that coordinator becomes a bottleneck: sub-agents wait on it, progress gets relayed through it, and peers redo work another peer already finished. The authors call this wasted time bubbles, agent time spent waiting on others or duplicating a peer’s result, and their answer is to remove the main agent entirely. For anyone running coding agents on client work, the pitch is unusually concrete: the same harness, minus the manager, measured on real long-horizon benchmarks. The question is no longer whether parallelism helps, but how much of it survives the coordination layer bolted on top of it.
What the paper does
The paper starts from how agents communicate: independent agents rediscover what peers already found, peer-communicating agents wait at synchronous rounds, and under centralized orchestration the main agent blocks on its sub-agents. Decentralized Language Models (DeLM) sits on top of existing agent harnesses and replaces the manager with a queue: agents asynchronously claim tasks, publish findings as soon as they are available, build on or correct one another’s progress, with every peer’s status visible to all. It is evaluated on Terminal-Bench 4.0, DeepSWE v1.1, and SWE-bench Verified against Codex, Claude Code, their native subagents, and AOrchestra, plus ProgramBench, where agents rebuild programs from scratch.
Findings
Where the waste hides — Parallelism is lost to idle time rather than compute, because agent time spent waiting on others or redoing a peer’s work never shows up as a wrong answer, only as a slower, costlier run. It never appears on a scoreboard, which is why token bills and accuracy tables both underreport it.
agent time spent waiting on others or redoing a peer’s work
The structural fix — DeLM does not tune the manager, it deletes it, and coordination moves into shared state that every agent can read and update. Progress becomes readable the moment it exists instead of waiting for a relay through the manager.
replacing the main agent with a shared context and a task queue
Measured payoff — Across every setting DeLM beats the harness it builds on, by up to 17.5 accuracy points over the strongest baseline and up to 2.49x in speed; on ProgramBench it finishes a 120-minute budget up to 19.9 points higher in test pass rate than Claude Code. The speedup and the accuracy gain arrive together, so this is not a quality-for-latency trade.
improving accuracy by up to 17.5 points over the strongest baseline and running up to 2.49x faster than the harness it builds on
Caveat
The authors state no limitation in the abstract. The evidence is bounded by the benchmarks named, Terminal-Bench 4.0, DeepSWE v1.1, SWE-bench Verified, and ProgramBench, and the abstract reports no failure cases for the shared-context queue, so its operating boundaries have to be read off those settings.
What this means for 5ac
The parent-child Kanban shape 5ac uses today is the centralized pattern this paper measures against: the orchestrator blocks while children run, and progress relays through it. The experiment to run is concrete: move a long-horizon coding card from parent-assigns-and-waits to a shared task queue with findings published as they land, then compare wall-clock time and cost per task against the parent-led version. Bubbles are the hidden line item clients never see on a token invoice, so any delegation quote should price coordination waiting, not just tokens spent by each profile.
Source
- Decentralized Multi-Agent Systems with Shared Context — https://arxiv.org/abs/2606.10662