Skip to content
Bud's Wiki

Decentralized Multi-Agent Systems with Shared Context

TL;DR

Section titled “TL;DR”

DeLM (Decentralized Language Models) replaces the central orchestrator in multi-agent LLM systems with a shared verified context + task queue: parallel agents asynchronously claim subtasks, read accumulated progress, perform local reasoning, and write back compact verified gists that become visible to all peers. On SWE-bench Verified, DeLM lifts Avg.@1 up to +9.3pp over the strongest centralized baseline (AOrchestra-Parallel) while cutting cost/task roughly in half (0.12vs0.12 vs 0.24 on Gemini 3 Flash). On LongBench-v2 Multi-Doc QA, DeLM gains up to +5.7pp average accuracy across four frontier model families. Central thesis: test-time scaling converts tokens into capability only when parallel agents share state, not when they share a controller.

Key claims

Section titled “Key claims”
  • Centralized orchestration (Claude Code Subagents, Kimi Agent Swarm, AOrchestra) is a two-sided bottleneck: it serializes progress-sharing through one main agent, and the main agent dilutes/omits/distorts useful details on the way through [§1]. DeLM moves coordination from prompt-routing to a shared state substrate [§2.1].
  • Shared context stores compact gists with hierarchical unfolding: agents read a lightweight global view by default and selectively expand into detailed summaries or raw evidence only when a subtask demands it [§2.2, §3.2]. Operating-system analogy: gists are the resident working set, detailed summaries + raw store are backing storage, unfolding is demand paging [§3.2].
  • Admission-time verification is load-bearing: every gist is checked against its underlying evidence (source + reasoning trajectory) before entering the shared context; post-hoc answer checking is insufficient because errors propagate as reusable state before any final answer is produced [§2.3, §3.2].
  • On SWE-bench Verified with Gemini 3 Flash, DeLM reaches 65.7% Avg.@1 / 72.9% Pass@2 / 77.4% Pass@4 at 0.12/task—strongestacrosseverymetricand halfthecostofAOrchestra/mini−SWE−agent/ClaudeCode[Table1].WithClaudeOpus4.6,DeLMreaches78.00.12/task — strongest across every metric and ~half the cost of AOrchestra / mini-SWE-agent / Claude Code [Table 1]. With Claude Opus 4.6, DeLM reaches 78.0% / 80.7% / 82.5% at 0.63/task, best on all three metrics [Table 1].
  • Trace-level mechanisms for the SWE-bench gain: (a) failed hypotheses become reusable state — one agent’s negative result (e.g. “changing the printer does not affect the output”) lets peers skip the same dead end; (b) constraints remain binding — AOrchestra-Parallel’s main agent softened a correctness constraint from a sub-agent into a tradeoff and the run failed, while DeLM kept the FAIL + predicate explicit in shared context; (c) compact patch summaries lower cost — the same successful solve costs 0.125vs0.125 vs 0.399 when the handoff is a compressed PATCH_SUMMARY rather than a raw transcript [§4.2.1].
  • On LongBench-v2 Multi-Doc QA, DeLM is best average across four model families: 60.1% (GPT-5.4, +5.7pp over best baseline), 59.8% (Claude Sonnet 4.6, +5.3pp), 61.5% (Gemini 3 Flash, +4.4pp), 67.5% (DeepSeek-V4-Pro, +3.6pp) [Table 2]. Gains attributed to upfront hierarchical summarization + verification building a reliable corpus-wide view before any targeted inspection; baselines (Claude Code, ReadAgent) make local inspection decisions from metadata and compound early misses [§4.3].
  • AOrchestra-Parallel (same AOrchestra orchestrator, but main agent spawns multiple sub-agents per step) improves Avg.@1 over vanilla AOrchestra but regresses on Pass@2 and Pass@4 — coupling parallel threads through a main agent increases consistency at the cost of exploration diversity [Table 1 / §4.2]. Direct evidence that “more agents” is not the same as “more parallel progress”.

Method

Section titled “Method”

Two global data structures: a shared context C (verified, admitted gists with pointers to detailed summaries and raw source) and a task queue Q (pending subtasks). Agents loop: claim a ready subtask from Q → read current C → reason locally → produce a candidate result → compress to a gist → verify the gist against evidence → admit on pass, reject/regenerate on fail. When Q is exhausted, the most recently completed agent decides from C whether to enqueue new subtasks or finalize the answer.

Compression has two paths. For a reasoning trajectory, DeLM directly summarizes into a gist (the trace is rarely useful downstream; only the distilled finding/failure/constraint is). For a long source unit, DeLM goes raw → reference-grounded summary S → compact gist G; both S and raw stay in backing stores and are recoverable through selective unfolding. Verification mirrors this: LLM verifier checks that reasoning gists faithfully capture the finding; for source gists, S must be supported by raw and G must preserve the supported claims and qualifiers from S.

On SWE-bench, DeLM uses the same agent harness as AOrchestra so the comparison isolates coordination substrate (shared verified context vs main-agent routing) rather than agent capability. Baselines include Base (single LLM call), Claude Code, mini-SWE-agent, AOrchestra, AOrchestra-Parallel (same AOrchestra but main agent spawns n subtasks per step), and ReadAgent (short gist memories with lookup).

Results

Section titled “Results”

SWE-bench Verified (Table 1). Gemini 3 Flash — DeLM 65.7 / 72.9 / 77.4 @ 0.12vsAOrchestra−Parallel56.4/63.2/71.8@0.12 vs AOrchestra-Parallel 56.4 / 63.2 / 71.8 @ 0.25 and mini-SWE-agent 54.7 / 65.6 / 75.1 @ 0.26.ClaudeOpus4.6—DeLM78.0/80.7/82.5@0.26. Claude Opus 4.6 — DeLM 78.0 / 80.7 / 82.5 @ 0.63 vs mini-SWE-agent 76.9 / 79.8 / 81.7 @ 0.61(DeLMbestonquality,within0.61 (DeLM best on quality, within 0.02 of lowest-cost baseline). AOrchestra-Parallel beats AOrchestra on Avg.@1 but loses on Pass@2/4 — centralized parallel coupling trades diversity for consistency.

LongBench-v2 Multi-Doc QA (Table 2). DeLM is top average across GPT-5.4 (60.1), Claude Sonnet 4.6 (59.8), Gemini 3 Flash (61.5), DeepSeek-V4-Pro (67.5), with per-domain wins or ties in most of {Financial, Government, MNews, Legal, Academic}. Gains attributed to the hierarchical-summary + verification upfront cost, which earns more targeted inspection downstream than ReadAgent’s lossy gists or Claude Code’s metadata-driven inspection.

OOLONG (§5, Table 3). Vanilla DeLM underperforms RLM on aggregation-heavy tasks where code-mediated execution is especially useful, but RLM + DeLM yields best accuracy and lowest cost — DeLM extends beyond conversational agents to compose with programmatic reasoning systems.

Why it’s interesting

Section titled “Why it’s interesting”

Direct contribution to Agentic Software Engineering: the “harness > model” thesis has been accumulating (Databricks’ per-task-cost finding; Qwen3-Coder-Next’s multi-scaffold training; SoL-Pi’s RSI-discovered harness) but DeLM isolates a specific harness lever — replace the controller with verified shared state — and shows it moves both Pass@k and $/task independently of base model (Gemini 3 Flash and Claude Opus 4.6 both benefit). The AOrchestra-Parallel ablation is the sharpest filed evidence that centralized parallelism (more sub-agents per step) is strictly inferior to decentralized parallelism (shared state) for SWE-bench-style exploration. Also complements PithTrain: A Compact and Agent-Native MoE Training System‘s ATE-Bench finding that codebase/harness design moves per-task cost by >2× at fixed model.

Within Parametric memory: shared verified context is a third form of agent-side memory beyond token-level and in-weights — Arjun Kumar’s f(log) sketch is the closest filed analog (append-only event log as substrate), but DeLM materializes it with admission-time verification and hierarchical unfolding as the compaction primitive, which the log-based sketch does not specify. Also complements Memory in the Age of AI Agents‘s Forms × Functions × Dynamics taxonomy: DeLM’s gist layer is token-level + working + formation-heavy, with verification as the Dynamics-axis gate.

See also

Section titled “See also”