AgentRoom: Concurrent Multi-Agent Coding in a CRDT-Backed Shared Workspace

Seonglae Cho; Donghyun Lee

cs.AIAug 24, 2026

AgentRoom: Concurrent Multi-Agent Coding in a CRDT-Backed Shared Workspace

Authors: Seonglae Cho, Donghyun Lee

Organizations: Holistic AI · University of California, Berkeley

Abstract

Concurrent multi-agent coding promises division of labor across modules, robustness through redundancy, and parallel exploration at the natural granularity of multi-file projects. Realtime collaborative editing protocols solve this coordination problem for human teams via Conflict-free Replicated Data Types (CRDTs), but the LLMs underneath generate one token at a time and existing multi-agent coding systems inherit this serial limit: they either sequence agents through phase handoffs or pool independent samples without coordination, and a single agent abandons up to half of hard tasks with a one-file stub-and-exit. AgentRoom is a realtime collaborative editing protocol for concurrent coding agents. Its runtime layer exposes file-level claim, status, and broadcast as MCP tools on a CRDT-merged shared filesystem. Five frontier coding-CLI models ran four backend coding tasks, with cross-language checks in Python DevBench and Rust+axum. For CLI-stable models, AgentRoom with 2 agents abandons fewer tasks than Solo and has less run-to-run variation. At matched-compute, one positive mean LLM-judge contrast puts AgentRoom over parallel-merge. The other contrast, a bundle probe, puts full AgentRoom above each partial case: an ordering rather than a percentage split. Coordination, not parallelism or CRDT-merge, bears the load.

Figures & tables

Figure 1: T4 six-condition coordination ablation: AgentRoom is the top condition, while concurrency without coordination (parallel-merge) and sequential hand-off (ChatDev-style) fall below the single-agent baseline.

Figure 2: AgentRoom architecture:

N

coding agents share a workspace mediated by a CRDT; the MCP server exposes claim , release , broadcast , state , read .

Figure 3: Real-time CRDT merge during a single T4 Sonnet AgentRoom

\times 2

run. The MCP-mediated chat (left) drives concurrent character-level edits in the shared file (right); both

\alpha

’s and

\beta

’s insertions co-exist on the same line via the Yrs CRDT, with no byte-level conflict raised in this run. Per-message Bezier leaders link each chat broadcast to the code line it talks about.

Figure 4: Solo (slate) vs AgentRoom

\times 2

(rust) score distributions for the three CLI-stable models on T4 (LLM-judge composite, budget-fair pool).

\sigma

contracts by

\sim 30

–

45

% under AgentRoom in every panel (Sonnet

0.23{\to}0.14

, Haiku

0.25{\to}0.17

, Codex

0.27{\to}0.15

); mean shift varies by model.

Condition	Mean	$n$	$\sigma$	Note
ChatDev-style (sequential 3-phase)	0.333	6	0.197	sequential paradigm baseline
Parallel merge (2 indep + union)	0.456	12	0.178	concurrent, no coordination
Solo (1 agent)	0.544	32	0.222	single-agent baseline
Shared only (2 agents, no MCP)	0.575	11	0.155	shared CRDT, no signalling
Shared + collab prompt, noMCP	0.588	7	0.188	+ prompt, no MCP tools
AgentRoom (2 agents, full)	0.669	14	0.140	CRDT + room MCP + collab prompt

Table 1: T4 Sonnet 4.6 six-condition ablation (LLM-judge composite, budget-fair

30

–

700

s pool). Parallel-merge vs AgentRoom:

+0.213

, Welch

t{=}3.35

\mathbf{p{=}0.003}

. The shared+collab-noMCP probe ran under a

1200

s wall budget; its runs are filtered to the same

30

–

700

s envelope as every other cell (Appendix B.13 ). Cross-model replication in Appendix B , Table 8 .

Figure 5: Agent-count scaling on T4 Sonnet 4.6 (

N{=}1

–

4

, budget-fair pool). Left: mean LLM-judge composite with

\sigma

band peaks at

N{=}2

then declines. Middle: mean tests passing peaks at

N{=}3

. Right: AgentRoom message count rises super-linearly past

N{=}3

, consistent with a single-broadcast-channel coordination overhead.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Cochran-Mantel-Haenszel forest plot of Solo-vs-AgentRoom

1

-file abandonment across all

12

model

\times

task strata (Sonnet/Haiku/Codex

\times

T1/T2/T4/T5).

Figure 7: Bootstrap effect-size distributions (

10\,000

resamples) for the Tier III cross-model cells. Only Haiku T4 has its

95\%

CI strictly above

1

; the other five cells span

0.79

–

1.96\times

point-estimate with CIs that include

1

Model	SWE-bench V.	Solo $(n)$	$\times$ 2 $(n)$	ES (AST / regex)	95% CI (AST)
Haiku 4.5	73.3%	0.520 (13)	0.929 (8)	1.79 $\times$ / 1.58 $\times$	[1.24, 3.35] ∗
Codex GPT-5.4	$\sim$ 80%	0.248 (9)	0.487 (13)	1.96 $\times$ / 1.07 $\times$	[0.85, 8.85]
Sonnet 4.6	79.6%	0.463 (11)	0.630 (15)	1.36 $\times$ / 1.14 $\times$	[0.84, 2.29]

Appendix

Table 2: Cross-model T4 ES with bootstrap 95% CI (10 000 resamples), AST-scored subsample of the full pool (

\geq 30

s elapsed).

n

is smaller than the LLM-scored pools because the AST scorer covers a subsample. Headline scorer is AST (TypeScript compiler API); regex-scorer values after the slash for cross-check. ∗ CI excludes 1. Two CLI-unstable models (Gemini, GPT-5.4-mini) excluded (see § 4.4 ). Reported as exploratory conjecture: only Haiku T4 clears the CI under the AST scorer.

Task	Solo (n)	$\times 2$ (n)	ES
T1	0.919 (4)	0.910 (7)	0.99 $\times$
T2	0.840 (3)	0.904 (4)	1.08 $\times$
T3	0.887 (3)	0.943 (4)	1.06 $\times$
T4	0.796 (10)	0.958 (10)	1.20 $\times$
T5	0.697 (7)	0.810 (7)	1.16 $\times$

Appendix

Table 3: Sonnet 4.6 ES across the difficulty gradient (regex/AST scorer, outlier-filtered

\geq 0.3

Pair	Score (n)	Tests
Sonnet $\times$ 2 (homo)	0.669 (14)	32.8
Sonnet+Codex	0.721 (3)	16.0
Sonnet+Gemini	0.542 (3)	22.0

Appendix

Table 4: Heterogeneous vs. homogeneous

\times 2

on T4 (LLM-judge, budget-fair). Tests is the mean over all budget-fair runs in the cell (

n{=}17

4

3

), which is not the LLM-scored

n

shown in the Score column.

Figure 8: Cost–quality Pareto on T4. Shaded region is Pareto-dominated; dotted lines are iso-efficiency (quality/cost) contours.

2{\times}

Haiku AgentRoom clears

1{\times}

Sonnet at matched plotted cost.

Condition	$N$	Wall (s)	Compute (s)	Quality
Solo	1	550	550	0.544
Parallel merge	2	558	1077	0.456
Shared only	2	601	1201	0.575
AgentRoom $\times$ 2	2	561	1072	0.669
AgentRoom $\times$ 3	3	566	1572	0.553

Appendix

Table 5: Wall vs compute on T4 Sonnet (LLM-judge composite, budget-fair pool).

Model	Solo $(n)$	$\times$ 2 $(n)$	$\times$ 3 $(n)$	ES( $\times$ 2)	Tests ( $\times$ 2)
Haiku 4.5	0.416 (8)	0.707 (7)	0.822 (1)	1.70 $\times$	38
Sonnet 4.6	0.618 (7)	0.674 (7)	0.730 (2)	1.09 $\times$	57
GPT-5.4	0.667 (9)	0.531 (7)	0.761 (1)	0.79 $\times$	8
Gemini Flash (unstable)	0.851 (3)	0.506 (9)	0.246 (3)	0.59 $\times$	12

Appendix

Table 6: T5 (trading platform, 900 s budget), AST scorer, per-model cells. All CLI-stable cells are point-estimate-positive on

\times 2

except Codex; CIs do not exclude 1 at this

n

. The Gemini row is CLI-unstable and informational only (§ 4.4 ); sparse heterogeneous cells are in the text above.

Behavior	Runs exhibiting (of 6)
Module claiming via broadcast	6/6
Plan adjustment after reading teammate	5/6
Export list shared for teammate	4/6
Proactive gap filling	3/6
Cross-agent bug fix	2/6
Apology for file boundary violation	2/6

Appendix

Table 7: Frequency of unprompted collaborative behaviors across

6

valid

\times 2

AgentRoom runs on T4 Sonnet.

Model	Solo	Parallel merge	Shared only	noMCP	AgentRoom $\times 2$
	Baselines and partial bundles				Ours
Sonnet 4.6	0.544 ( $n{=}32$ )	0.456 ( $n{=}12$ )	0.575 ( $n{=}11$ )	0.588 ( $n{=}7$ )	0.669 ( $n{=}14$ )
Haiku 4.5	0.530 ( $n{=}35$ )	0.383 ( $n{=}12$ )	0.657 ( $n{=}7$ )	0.579 ( $n{=}6$ )	0.662 ( $n{=}17$ )
Codex GPT-5.4	0.384 ( $n{=}12$ )	–	–	–	0.529 ( $n{=}17$ )

Appendix

Table 8: Cross-model T4 ablation (LLM-judge composite, budget-fair pool). For Sonnet and Haiku, AgentRoom is the top condition and parallel-merge falls below solo. Codex non-Solo/AgentRoom cells were not collected (vendor-CLI hang).

Thresholds		Sonnet 4.6		Haiku 4.5		Codex GPT-5.4		CMH pooled
$d$ thr	$f$ thr	Solo	Room	Solo	Room	Solo	Room	OR	95% CI	$p$
$<150$	$\leq 1$	6/33	0/17	12/35	1/17	9/17	0/21	24.2	[3.5, 169]	0.00001
$<150$	$\leq 2$	6/33	0/17	12/35	1/17	9/17	2/21	11.9	[3.0, 47]	0.00005
$<150$	$\leq 3$	6/33	0/17	12/35	1/17	10/17	6/21	6.3	[2.1, 19]	0.00069
$<200$	$\leq 1$	6/33	0/17	12/35	1/17	9/17	0/21	24.2	[3.5, 169]	0.00001
$<200$	$\leq 2$ (paper)	6/33	0/17	12/35	1/17	9/17	2/21	11.9	[3.0, 47]	0.00005

Appendix

Table 9: Threshold sensitivity for the Tier I CMH. Each row is a threshold combination; cells report Solo/AgentRoom abandonment counts per model and the resulting CMH common OR with

95\%

CI and

p

-value.

Pool	Solo	Parallel merge	Shared only	AgentRoom $\times 2$
$30$ – $700$ s (paper, budget-fair)	0.544 ( $n{=}32$ )	0.456 ( $n{=}12$ )	0.575 ( $n{=}11$ )	0.669 ( $n{=}14$ )
$\geq 30$ s (full pool)	0.464 ( $n{=}42$ )	0.436 ( $n{=}13$ )	0.527 ( $n{=}14$ )	0.639 ( $n{=}15$ )

Appendix

Table 10: Pool sensitivity for the T4 Sonnet ablation (LLM-judge composite). The substrate-vs-AgentRoom ordering is preserved under both pools; the budget-fair pool is primary.

Model	Solo (bare, budget-fair)	Solo + collab prompt ( $n{=}3$ )	AgentRoom $\times 2$ (budget-fair)
Sonnet 4.6	0.544 ( $n{=}32$ )	0.448 (range 0.20–0.62)	0.669 ( $n{=}14$ )
Haiku 4.5	0.530 ( $n{=}35$ )	0.604 (range 0.55–0.66)	0.662 ( $n{=}17$ )

Appendix

Table 11: Solo + collaboration-style prompt without teammates, T4,

n{=}3

each (LLM-judge composite). The structured prompt alone is model-dependent: it hurts Sonnet (channel is the load-bearing variable) and helps Haiku (the structured prompt also acts as workflow scaffolding); the channel and prompt-as-workflow are not separately identified for Haiku at this

n

Task	Solo (PASS, dur)	$\times$ 2 AgentRoom (PASS, dur)
arxiv-digest	PASS, 70 s	PASS, 120 s
readtime	PASS, 110 s	PASS, 100 s
chakin	PASS, 130 s	PASS, 130 s
geotext	FAIL, 603 s timeout	PASS, 390 s
hone	PASS, 90 s	PASS, 160 s
hybrid-images	PASS, 100 s	PASS, 170 s

Appendix

Table 12: DevBench cross-domain validation (Haiku 4.5, 600 s budget). Pass = hidden pytest suite passes; durations in seconds. geotext :

\times 2

completes within budget while solo times out; stocktrends, textcnn : both fail (task too hard for budget).

Figure 9: Single representative T4 Sonnet AgentRoom

\times 2

run. Top swim-lane: agent

\alpha

broadcasts (timestamps in-card); middle: per-file ownership Gantt (rust =

\alpha

, slate =

\beta

); bottom: agent

\beta

. Bezier S-curve leaders link each chat broadcast to its time on the Gantt. The whole protocol (claim, conflict-detect, hand-off, completion broadcast) emerges from the advisory six-step prompt (§ 2.3 ) without any orchestrator.

Figure 10: The two scorers measure the same thing loosely, not interchangeably. Each point is one T4 run carrying both a regex composite and an LLM-judge composite (

n{=}246

, the runs scored by both); Pearson

r{=}0.67

. The spread off the diagonal is why the body uses one scorer throughout and reports the other as a cross-check rather than averaging them.

Removed dimension	Solo mean	$\times$ 2 mean	ES
(none, full scorer)	0.786	0.928	1.18
Type Safety	0.721	0.904	1.25
Defensive Coding	0.815	0.944	1.16
Test Quality	0.854	0.956	1.12
Design	0.793	0.921	1.16
Test Pass Rate	0.749	0.916	1.22

Appendix

Table 13: Scorer sensitivity: leave-one-out on T4 (Sonnet 4.6, solo vs

\times

2 AgentRoom). The solo mean reported here (0.786) is computed over a smaller subset (4 runs) than the canonical solo subset used in Table 1 ; the smaller subset is the intersection of solo runs for which we retained per-dimension intermediate scorer outputs at the time of writing, kept here only as a robustness check. The AgentRoom

>

solo ordering is preserved under every single-dimension ablation and the ratio stays within 1.12–1.25 on this subset.

Explore similar work

AgentRadio: Passive Awareness for Long-Horizon Multi-Agent Collaboration
Jul 30, 2026Xinxing Ren, Qianbo Zang, Ziyan Wang +4Software Engineering Agents Long-Horizon LLM Agents
1Coral AI Labs · 2SnT, Université du Luxembourg · 3King’s College London +1
OpenCollab: A Multi-Agent Coding Framework with Programmable Collaboration and Controllable Runtime
Sep 29, 2026Chun-Wah Hsu, Kai Gong, Yu Wu +12Multi-Agent Orchestration Multi-Agent Collaboration
Shanghai Jiao Tong University · University of Cambridge · Nanyang Technological University +4
When Parallelism Pays Off: Cohesion-Aware Task Partitioning for Multi-Agent Coding
May 31, 2026Xu Yang, Lunyiu Nie, Ethan Chandra +3Multi-Agent System Optimization Software Engineering Agents
1The University of Texas at Austin · University of Oxford

Jul 30, 2026J/K move · Enter open · S save

AgentRadio: Passive Awareness for Long-Horizon Multi-Agent Collaboration

Xinxing Ren, Qianbo Zang, Ziyan Wang, Caelum Forder+3

1Coral AI Labs · 2SnT, Université du Luxembourg · 3King’s College London+1

Software Engineering Agents Long-Horizon LLM Agents Multi-Agent Communication Multi-Agent Coding

Open paper AskPDF arXiv

Understanding large codebases is a long-horizon task for Large Language Model (LLM) agents: answering a single question can require building and running the software, tracing execution across files, and synthesizing evidence over tens of minutes. On SWE-Atlas QnA, a benchmark of long-horizon questions over production repositories, a single Claude Code agent (Opus 4.6) resolves only 32.3% of tasks. Dividing the work among agents with clean contexts mitigates this limitation. However, the subtasks of code comprehension are interdependent. One agent's findings can rewrite another's task, so agents must coordinate during execution, not only at phase boundaries. Existing multi-agent systems support such exchange only between phases, through staged handoffs or synchronized rounds. Communication and work remain mutually exclusive. A discovery made mid-execution cannot be shared until the next boundary. We present AgentRadio, an asynchronous message-passing layer that equips coding-agent harnesses with three primitives: threads, messages, and waiting for mentions. The last runs as a background task, surfacing teammates' messages without interrupting foreground work, so each agent remains passively aware of its peers and folds new findings into its ongoing task. Under a five-phase protocol of division of labor and negotiation, four agents organized by AgentRadio resolve 62.1% of tasks, 29.8 points above a single agent and above Claude Code with the newer Opus 4.8 (57.2%). Rubric-level analysis shows the gain growing with task difficulty, consistent with mid-course correction as the underlying mechanism. Our code is available at https://github.com/Coral-Protocol/AgentRadio.

Abstract

Explore similar work

AgentRadio: Passive Awareness for Long-Horizon Multi-Agent Collaboration

OpenCollab: A Multi-Agent Coding Framework with Programmable Collaboration and Controllable Runtime

When Parallelism Pays Off: Cohesion-Aware Task Partitioning for Multi-Agent Coding