The Big Picture
Shared, verified problem state lets many AI agents coordinate without a central controller, improving accuracy on long-context reasoning and cutting test-time cost—e.g., 77.4% pass@4 on coding tasks at roughly half the cost of centralized baselines.
ON THIS PAGE
The Evidence
A shared, verified context plus a decentralized task queue lets multiple agents read and build on compact "gists" of prior work instead of routing everything through one master agent. On a software-engineering benchmark, this approach turned parallel attempts into reusable exploration, boosting pass rates while halving cost. On multi-document question answering, the shared context and hierarchical summaries raised average accuracy by several points. For tasks that need exact programmatic aggregation, combining the decentralized coordination with code-based execution produced the best accuracy and lowest cost. Agentic RAG Pattern
Data Highlights
177.4% pass@4 on SWE-bench Verified when using the decentralized shared-context system
2$0.12 average cost per task on SWE-bench Verified — roughly half the cost of centralized baselines
3Up to 5.7 percentage points higher average accuracy on LongBench-v2 multi-document QA versus the best baseline
What This Means
Engineers building agentic systems and toolchains: adopt a shared verified state to make parallel agent runs reuse discoveries and avoid repeated failures. Technical leaders evaluating agent orchestration: decentralized coordination can cut test-time costs and raise QA performance without adding a central bottleneck. Researchers designing large-scale workflows: the approach scales better for long-context reasoning and is a natural fit for automated research pipelines. See also Multi-Agent Knowledge Management
Not sure where to start?Get personalized recommendations
Key Figures
Fig 1: Figure 1: Comparison on SWE-bench Verified and LongBench-v2 Multi-Doc QA. Our method, DeLM , achieves the best average performance across both agentic and long-context benchmarks.
Fig 2: Figure 2: Centralized vs. decentralized multi-agent systems. Centralized MAS relies on a main agent to assign subcontexts, spawn sub-agents, and integrate returned results through a synchronous scatter–gather loop, creating a bottleneck where progress is gated by both the central merge step and the slowest worker. In contrast, DeLM decentralizes coordination through parallel agents, a shared context, and a task queue, allowing agents to exchange progress through shared state, asynchronously claim ready tasks, and scale more adaptively as the number of subtasks grows.
Fig 3: Figure 3: Overview of DeLM . A one-time initialization step decomposes the input into initial subtasks and places them in a shared task queue. Parallel agents asynchronously claim tasks ( T i {T_{i}} ), read the verified shared context, and perform local reasoning. Completed updates are compressed, verified, and admitted as compact gists ( G i {G_{i}} ), making reusable progress visible to all agents. When the task queue becomes empty, the most recently completed agent checks the existing subtask states and shared context to decide whether additional subtasks are needed. If so, it generates and enqueues new tasks for another round of parallel execution; otherwise, it produces the final answer.
Fig 4: Figure 4: Ablation and robustness analysis on LongBench-v2 Multi-Doc QA. All bars report accuracy averaged over the five domains with GPT-5.4. (a) Modular ablation: removing either the verification step or the hierarchical summary lowers accuracy. (b, c) DeLM is largely insensitive to its gist configuration: accuracy is stable once the gist is long enough (b) and across summarizers of varying cost (c). In every panel, the darker bar marks the default configuration used in Table 2 .
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
Admitting only verified updates improves reliability but adds runtime overhead; lighter or learned verifiers could reduce that cost. Performance depends on how tasks are split—too coarse or too fine decomposition hurts coordination and efficiency. For structured aggregation tasks (counts, exact filters), natural-language shared state is less reliable, so combining decentralized coordination with programmatic execution may be necessary. Sub-Agent Delegation Pattern
Methodology & More
DeLM replaces a single orchestrator with two shared structures: a task queue and a verified shared context that stores compact "gists" of progress. Agents asynchronously claim subtasks, consult the shared gists, perform local reasoning, and propose compressed updates. Each update must pass an admission-time verification step before becoming visible to others; admitted gists can be unfolded back to summaries or raw evidence on demand. That state-based communication avoids the serial merge-and-redistribute bottleneck of centralized systems and keeps the global view lightweight and grounded in evidence. Market-Based Coordination Pattern Hallucination Propagation
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Includes a well-established researcher (Azalia Mirhoseini, h-index 36) indicating strong credibility, though listed as an arXiv preprint.