LLM-based multi-agent systems (MAS) rely on collaboration among specialized agents to solve complex tasks. Existing methods improve MAS by refining the prompts, skills, and execution strategies of individual agents based on task feedback. However, improving individual agents does not necessarily lead to better team collaboration. Recent studies have therefore extended evolution to the collaboration level, using collective execution experience to adjust team organization, and inter-agent interactions. Nevertheless, evolving collaboration across heterogeneous tasks remains challenging, as existing approaches often constrain team reorganization within a shared structure and reuse interaction rules across incompatible team contexts. In this paper, we propose EvoTeam, a collaboration-centric evolution framework that jointly evolves team organization, interaction rules, and individual agent skills through execution experience. It maintains a library of reusable teams, selecting or creating agent teams and guiding collaboration through team-specific interaction rules. Collaboration experience accumulates within compatible team contexts, while individual skills remain reusable across teams. Experiments across four benchmarks demonstrate its effectiveness. Our code is available at https://github.com/DEEP-JLU/EvoTeam.
Figures & tables
Figure 1: Three paradigms of MAS.
Figure 2: Preliminary study: experience reuse in MAS. (a) Cross-task training underperforms target-only training on target tasks. (b) Trajectories show how inflexible templates and unclear rule scopes hinder reuse.
Figure 3: The overall framework of EvoTeam . (a) Adaptive Agent Team Composition. The system selects or creates a team, and the Chairman recruits agents from its candidate roster. (b) Team-Scoped Collaboration. Agents execute assigned subtasks using accumulated skills and applicable team-specific rules. (c) Multi-Level Experience Evolution. Reflectors refine teams, team-specific rules, and agent skills using execution traces.
Method
GAIA
LoCoBench
LOCA
BeyondSWE
Avg.
Val
Feat.
Refact.
Val
DepMig.
CrossR.
Individual-Benchmark Evolution
SA
64.00
55.99
53.08
37.50
34.81
37.22
47.10
MAS
72.00 ↑8.00
61.71 ↑5.72
61.49 ↑8.41
49.17 ↑11.67
29.75 ↓5.06
38.33 ↑1.11
52.08 ↑4.98
OWL
48.00 ↓16.00
51.80 ↓4.19
50.64 ↓2.44
39.17 ↑1.67
33.54 ↓1.27
39.44 ↑2.22
43.77 ↓3.33
AgentSquare
49.33 ↓14.67
62.45 ↑6.46
60.55 ↑7.47
43.33 ↑5.83
34.81 —
35.56 ↓1.66
47.67 ↑0.57
Table 1: Main results across diverse agent benchmarks under both individual-benchmark and mixed-benchmark evolution. Subscripts show absolute score differences from SA. The best scores within each block are in bold .
Evolution → Evaluation
GAIA
LoCoBench Feat
LoCoBench Refact
MAS
Ours
Δ
MAS
Ours
Δ
MAS
Ours
Δ
Same-Backbone Evaluation
DeepSeek → DeepSeek
58.00
80.00
+22.00
66.32
68.93
+2.61
65.15
68.16
+3.01
GPT → GPT
72.00
78.00
+6.00
61.71
67.32
+5.61
61.49
66.23
+4.74
Cross-Backbone Evaluation
DeepSeek → GPT
72.00
77.00
+5.00
61.71
65.28
+3.57
61.49
66.15
+4.66
Table 2: Same-backbone and cross-backbone evaluation. DeepSeek and GPT denote DeepSeek-V4.1-Flash and GPT-5.6-Luna, respectively. Arrows indicate the evolution-to-evaluation direction. Bold indicates the better score.
Figure 6
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Task
Holdout
Domain
Key MCP services
WoocommerceNewWelcome
30
E-commerce CRM
woocommerce, google_cloud, email
WoocommerceStockAlert
30
E-commerce purchasing
woocommerce, google_sheet, email
FilterLowSellingProducts
30
E-commerce clearance
woocommerce, email
ApplyPhDEmail
30
Academic
email, pdf_tools
SetConfCrDdl
30
Academic
calendar, email
CourseAssistant
30
Education
excel, email
Appendix
Table 3: The eight LOCA-Bench task types used in our experiments, following Meta-Team ( Hao et al., 2026c ) . Each task type has a total of 30 distinct holdout cases (five seeds across six environment-description lengths).
Benchmark / track
Evolution
Holdout
Native environment
Metric
GAIA
20
100
Attachments and web tools
Pass rate
LoCoBench FI
20
80
Isolated code workspace
LCBS (0–100)
LoCoBench CR
20
80
Isolated code workspace
LCBS (0–100)
LOCA-Bench
16
240
Local mock MCP services
Pass rate
BeyondSWE CrossRepo
20
180
Isolated repository
Pass rate
BeyondSWE DepMigrate
20
158
Isolated repository
Pass rate
Appendix
Table 4: Data splits and evaluation protocols. Counts denote distinct instances or environment configurations.
Benchmark
Max. time (s)
Max. messages
Max. cost ($)
GAIA
900
300
2.40
LoCoBench
1800
400
2.40
BeyondSWE
1800
600
2.40
LOCA-Bench
3200
600
12.00
Appendix
Table 5: Specified per-case resource budgets for both evolution and holdout evaluation phases.
Figure 6: Ablation results across six different evaluation metrics. We compare the proposed full model with variants that remove team-specific handoff rules, or both reusable agent teams and handoff rules.
Subset
Method
SE
FC
CQ
LCU
Feat.
Skill-MAS
0.4916
0.5744
0.5878
0.7360
ReCreate
0.3766
0.6832
0.8407
0.7384
Meta-Team
0.4923
0.7231
0.7802
0.7915
Ours
0.4979
0.7239
0.8064
0.7789
Refact.
Skill-MAS
0.5109
0.6264
0.5960
0.6686
ReCreate
0.4375
0.7181
0.7676
0.7001
Appendix
Table 6: LoCoBench dimension scores after mixed-benchmark evolution. The highest score for each dimension within each subset is in bold .
Level
Skill-MAS
Meta-Team
ReCreate
Ours
Level 1
42.11
78.95
69.30
81.58
Level 2
32.65
67.35
65.31
83.67
Level 3
30.77
30.77
30.77
38.46
Appendix
Table 7: GAIA pass rates (%) by difficulty level. Scores use all instances in each level. The highest score in each row is in bold .
Figure 8: Fine-grained performance on the LoCoBench Feature Implementation (Feat.) and Cross-File Refactoring (Refact.) subsets. Higher is better.
Figure 9: Prompt for team selection.
Figure 11: Prompt for Chairman coordination.
Figure 13: Prompt for team reflection.
Figure 15: Prompt for agent reflection.
Figure 16: Family-scoped experience accumulation versus shared-profile contamination on CourseAssistant. (a) Meta-Team replaces accumulated shared-profile notes with task-derived advice, removing earlier guidance and introducing an unscoped recruitment restriction. Strikethroughs indicate overwritten guidance; crosses mark problematic replacements. (b) EvoTeam preserves existing rules, appends supporting evidence, and adds a verification rule within the selected family. All 198 rules in the other six families remain unchanged.
Figure 17: Cross-benchmark collaboration evolution. (a) Team creation and reuse. (b) Collaboration guided by team-scoped interaction rules. (c) Accumulated rule updates and candidate-roster expansion across four tasks.
Table 8: Simplified trace of the Case 3 escrow development task (LoCoBench).
Figure 20: Team configuration and rule-guided collaboration on the Carnival refactoring task. (a) Our method recruits an independent verifier and attaches team-scoped rules to the initial handoffs. Meta-Team recruits the analyst and implementer without a separate verifier. (b) Our team corrects an interface mismatch and independently rechecks the revised files. Meta-Team’s later edits lack updated validation, leaving an invalid wrapper in the submission. (c) Final LCBS scores (out of 100).