Team Sizes#

Each multi-agent topology also runs with 2, 4, 8 and 10 agents. This page shows what grows with the team in each topology, and how to run and optimize every size.

Sizesn = 2, 4, 8, 10

Topologies4

Datasets9

Runners144

What n changes#

Single is excluded: one agent has no team size. For the other four topologies, n = 4 is the baseline: the team of the LangGraph runners in topologies/.

Topology	What n sets	n = 2	n = 4	n = 8	n = 10
Independent	parallel replicas	2	4	8	10
Sequential	pipeline stages	2	4	8	10
Centralized	1 manager + (n − 1) workers	1 + 1	1 + 3	1 + 7	1 + 9
Decentralized	debating peers, always 2 rounds	2 × 2	4 × 2	8 × 2	10 × 2

For seven datasets, every size of the team is declared in the team spec, configs/teams/<dataset>.yaml; ToolHop and API-Bank work differently (see below). Tools, aggregation rules and the Decentralized round count stay as in the baseline. What grows is the set of agents:

Independent and Decentralized add copies of the one shared prompt (solver, coder, caller or patcher; debater).
Sequential and Centralized add new roles. At n = 2 the team keeps the essential tool-using role and the role that writes the answer. At n = 8 and n = 10 it adds specialist roles defined in configs/prompts/roles.yaml, whose seed prompts sit beside the baseline ones in configs/prompts/<topology>/<dataset>/.
Centralized swaps in manager_r8.txt or manager_r10.txt at n = 8 and n = 10; those prompts name all 7 or 9 workers. At n = 2 the manager's turn cap is halved (18 to 9 on HotpotQA).

For example, the HotpotQA Sequential pipeline at each size:

n	Stages, in order
2	`retriever`, `writer`
4	`planner`, `retriever`, `reasoner`, `writer`
8	`query_decomposer`, `planner`, `searcher`, `retriever`, `evidence_filter`, `reasoner`, `citation_compiler`, `writer`
10	the 8 above plus `entity_disambiguator` (after `searcher`) and `answer_simplifier` (last)

Shell variables override n

Apart from ToolHop and API-Bank, the Independent and Decentralized runners read their size from INDEPENDENT_N_AGENTS or DECENTRALIZED_N_AGENTS first and only fall back to the n of the team spec. Unset both before running team-size cells.

ToolHop and API-Bank

These two datasets have no team spec. For them every team-size runner goes through a shared wrapper (teamsizes/toolhop_common.py, teamsizes/apibank_common.py). It runs n seeded copies of the topology's answering role (solver, verifier, manager or debater) and majority-votes their answers, for all four topologies. The optimizer adapters for these team sizes run the real topology at size n instead.

Files#

Runners follow teamsizes/<topology>/<dataset>/<dataset>_r<N>.py, for every topology, all nine datasets and N in 2, 4, 8, 10; for example, teamsizes/centralized/hotpotqa/hotpotqa_r8.py. Each file is a few lines: it runs the topology's LangGraph runner from topologies/ with the size-N team preset (core/variant.py), or the ToolHop and API-Bank wrapper.

Run a cell#

Run from the repository root. Team-size runners take the same command line as the topology runners. Their default output folders differ by runner, and some only print scores, so pass --out-dir:

GPQA, HotpotQA, MATH, LCB, APPSBFCLToolHop, API-Bank

python -m teamsizes.centralized.hotpotqa.hotpotqa_r8 --batch --limit 100 \
  --out-dir results/teamsizes/r8/centralized_hotpotqa

# always a batch, one category at a time
python -m teamsizes.sequential.bfcl.bfcl_r4 --category simple --limit 100 \
  --out-dir results/teamsizes/r4/sequential_bfcl/simple

export TOOLHOP_ALLOW_DATASET_EXEC=1
python -m teamsizes.decentralized.toolhop.toolhop_r10 --limit 100 \
  --out-dir results/teamsizes/r10/decentralized_toolhop

Without --batch, the GPQA, HotpotQA, MATH, LiveCodeBench and APPS runners play a built-in smoke demo. SWE-bench cells take the SWE-bench options.

Run the sweep#

scripts/run_teamsizes.sh loops over RVALUES, TOPOLOGIES and DATASETS (all values by default) and runs each cell as its own process:

Sweep two sizes of Centralized on HotpotQA

RVALUES="2 10" TOPOLOGIES="centralized" DATASETS="hotpotqa" \
  bash scripts/run_teamsizes.sh

It writes each cell to results/teamsizes/r<N>/<topology>_<dataset>/ (OUT_ROOT changes the root; BFCL gets one subfolder per category). BFCL and SWE-bench run their eval IDs; the other datasets use the same limits as the topology sweep.

Optimize a size#

Pass --team-size N to the run protocol, or use the key <topology>_r<N>. The bridge has team-size adapters for HotpotQA, LiveCodeBench, BFCL, ToolHop and API-Bank; the experiment grid covers HotpotQA, LiveCodeBench and BFCL with GEPA, MIPRO, MAPRO and MASPO, and other cells need --allow-any-cell. Decentralized jobs keep 2 rounds at every size. For example, MIPRO on a Centralized team of 8:

MIPRO on Centralized, 8 agents · HotpotQA

python -m optimizers.protocol.run --method mipro --dataset hotpotqa \
  --topology centralized --team-size 8 \
  --model qwen --seed 0 --out runs/mipro/hotpotqa/centralized_r8/qwen/0

See Run an Optimizer.