Large language model agents must discover relevant tools from large libraries, e.g., thousands of tools, as new needs arise during task execution. When several candidates could satisfy the same need, execution may be required to identify the relevant one. Investigating candidates in a shared trajectory can leave alternatives unexplored or allow failed attempts to influence subsequent reasoning. To address this, we present BRIA (BRanch-Inspect-Adopt), a simple yet effective agent that temporarily isolates candidate investigations in separate branches. Each branch retrieves and executes tools to investigate its candidate; a task agent inspects the resulting execution evidence and adopts only useful results into its continuing trajectory. BRIA retrieves additional tools as new capability needs arise during task execution. We evaluate BRIA on PlanBench-XL and MCP-Atlas using a shared pool of 1897 tools, three language models, and static and adaptive retrieval baselines. Across all settings, BRIA matches or exceeds the strongest retrieval baseline. On MCP-Atlas, it achieves an average claim score of 89.77, compared with 75.40 for MCP-Zero, a gain of 14.37 percentage points. BRIA also lies on the Pareto frontier for task performance and input-token cost.
Figures & tables
Figure 1: (a) Static top- k retrieval may miss the right source. (b) Adaptive retrieval tries sources sequentially in one trajectory: it may stop after the first failure (①) or accumulate failed attempts in context (②). (c) BRIA explores each source in an isolated branch and adopts only the one grounded in execution evidence ( e ).
Method
Iterative Retrieval
Branching
Selective Adoption
Plug-and-Play
One-shot Top- k
✗
✗
✗
✓
Tree-Planner ( Hu et al., 2024 )
✗
✓
✗
✓
MCP-Zero ( Fei et al., 2025 )
✓
✗
✗
✓
FitText ( Zheng et al., 2026 )
✓
✓
✗
✓
ToolGate ( Liu et al., 2026c )
✓
✗
✓
✓
ToolQP ( Fang and Glass, 2026 )
✓
✗
✗
✗
Table 1: Comparison of adaptive retrieval approaches. “Branching” denotes explicit investigation of candidate tool-use trajectories; “Selective adoption” denotes tool execution and selective retention of results from candidate investigations; and “Plug-and-Play” indicates whether an arbitrary LLM can be used without parameter updates.
Figure 2: The BRIA workflow. The task agent opens a checkpoint when ∣C∣≥2 candidates need the same investigation. ① Each candidate is investigated in an isolated branch (G1). ② The task agent compares branch evidence and adopts only useful branches (G2). ③ It resumes from the adopted context, retrieving tools (+T) whenever a capability is missing (G3), and opens the next checkpoint when another capability need has ∣C∣≥2 candidates.
PlanBench-XL
MCP-Atlas
Method
GPT-6 Luna
Gemini-3.8 Flash
GLM-5.3 Flash
GPT-6 Luna
Gemini-3.8 Flash
GLM-5.3 Flash
Task-Gold-Only
76.00
100.00
100.00
66.48
72.15
78.14
One-shot k=50
0.00
0.00
0.00
48.27
56.64
64.10
One-shot k=250
8.00
10.67
10.67
54.51
59.93
66.71
One-shot k=500
18.67
30.67
22.67
60.39
59.07
71.50
MCP-Zero
18.67
37.33
32.00
60.47
75.40
78.67
Table 2: Overall scores. PlanBench-XL: answer-correctness ∧ solution-path-coverage; MCP-Atlas: average claim score. Bold numbers indicate the best performance.
Figure 3: Comparison of sampled and full-set scores with GPT-6-Luna.
Figure 4: Score vs total input token amount.
Method
GPT-6 Luna
Gemini-3.8 Flash
GLM-5.3 Flash
Task-Gold-only
76.00
100.00
100.00
Gold + Blockers
48.00
96.00
92.00
BRIA w/o Retrieval
88.00
98.67
97.33
Table 3: No-retrieval setup comparison on PlanBench-XL. Gold-only and Gold + Blockers use a standard LLM agent; BRIA w/o Retrieval uses the Gold + Blockers tool menu.
Model
Method
Score (%)
Input (M)
PlanBench-XL (Success Rate)
GPT-6-Luna
BRIA
22.67
49.45
No Inspection
24.00
52.20
Gemini-3.8-Flash
BRIA
37.33
76.23
No Inspection
37.33
79.29
GLM-5.3-Flash
BRIA
38.67
66.43
Table 4: BRIA vs. No Inspection on 75 tasks per benchmark. “Input” denotes the total input token amount.
Method
PlanBench-XL Success (%)
MCP-Atlas Claim Score (%)
BRIA
22.67
62.35
No Repeat-Query Novelty
16.00
60.51
Table 5: Repeated-query novelty ablation with GPT-6-Luna on 75 tasks per benchmark.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
Value
Embedding model
Qwen3-Embedding-8B
Retrieval top- k
50
Max retrievals / task
20
Max tool calls / task
100
Max model calls / task
200
Max internal actions / task
200
Appendix
Table 6: Final BRIA configuration. All budgets are shared across the task agent and all branches of a task.
Model
Size
Context
Model name / snapshot
License
GPT-6-Luna
—
1M
gpt-6-luna
OpenAI Service Terms 3 3 3 https://openai.com/policies/services-agreement/ [Accessed: October 2, 2026]
Gemini-3.8-Flash
—
1M
gemini-3.8-flash
Gemini API Additional Terms of Service 4 4 4 https://ai.google.dev/gemini-api/terms [Accessed: October 2, 2026]
GLM-5.3-Flash
320B
300K
zai-org/GLM-5.3-Flash
MIT License
Appendix
Table 7: A list of the LLMs used in the experiments. An em-dash (—) means that the model size is not publicly disclosed.