Single#

One agent solves the task end to end in a reason-act loop with its tools. It is the control condition: every multi-agent result is read against it.

How it works#

1. The runner reads the seed prompt configs/prompts/single/<dataset>/solver.txt and wraps it with the dataset's protected output contract from core/output_contracts.py (for HotpotQA: end with one line Answer: <short-form>). HotpotQA and MATH also append a short format note.
2. It builds one agent with LangGraph's prebuilt create_react_agent, bound to the dataset's tools: wikipedia_search and wikipedia_page for HotpotQA, calculator for GPQA and MATH, python_exec for LiveCodeBench and APPS, file_read, file_write, list_dir, search_repo and shell_exec for SWE-bench, and tools built from the function schemas for BFCL.
3. The agent alternates between reasoning and tool calls. Each tool result is appended to its history. The loop ends when the model replies without a tool call, or when LangGraph's recursion limit is reached (25 steps; 100 for SWE-bench).
4. The runner parses the final message with the dataset's extractor and scores it.

There is no hand-off, no second agent and no vote. Whatever the one prompt makes the model do is the result.

Agent role and seed prompt#

Every dataset uses a single role, solver:

The role descriptions the prompts were generated from are under single: in configs/prompts/roles.yaml. The optimizers read these files and never overwrite them; a job keeps its prompts under its own --out folder.

Implementation#

• Reference scaffold: topologies/single/langgraph_base.py shows the pattern in 54 lines: create_react_agent with toy tools (add, multiply, a mock web_search). It is a demo, not a benchmark runner.
• Dataset runners: topologies/single/<dataset>/langgraph_<dataset>.py, one per dataset. Each wires the real tools to the same ReAct agent; the loader, scorer and records come from the dataset's task module in core/tasks/.
• ToolHop and API-Bank: these two runners don't use LangGraph objects. ToolHop runs a plain tool loop on the openai client (9 turns by default, TOOLHOP_MAX_TURNS); API-Bank makes one model call. Both live in core/tasks/toolhop.py and core/tasks/apibank.py, shared by every topology's runner.

Run it#

Run from the repository root with a model endpoint configured (Connect a Model).

python -m topologies.single.hotpotqa.langgraph_hotpotqa

python -m topologies.single.hotpotqa.langgraph_hotpotqa --batch --limit 100 \
  --out-dir results/topologies_baseline/single_hotpotqa

Other datasets follow the same module path with a different folder and file name, for example topologies.single.math.langgraph_math. BFCL, SWE-bench, ToolHop and API-Bank have no demo and always run a batch. See the task pages for each dataset's options.

Optimize it#

The protocol topology is single, with one prompt to tune, solver. In the experiment grid, the Single topology is optimized by GEPA and MIPRO; the other six methods optimize multi-agent teams.

python -m optimizers.protocol.run --method gepa --dataset math --topology single \
  --model qwen --seed 0 --out runs/gepa/math/single/qwen/0

See Run an Optimizer for endpoints and outputs.