AFlow — Agentic workflow search¶

Workflow self-evolution. Search the space of code-expressed agentic workflows, with soft mixed selection over the top-k scored ones. Runs through the shared MethodPolicy runner. Example: examples/aflow/aflow_workflow_search.py.


Paper	AFlow: Automating Agentic Workflow Generation — Zhang et al., ICLR 2025 (arXiv:2410.10762)
Upstream code	FoundationAgents/AFlow@3f457218
Example	`examples/aflow/aflow_workflow_search.py`
Domain	GSM8K (`openai/gsm8k`, `main`), 64/64/64 splits
Layer	L1 (`blast_radius=0.6`, set by the shared runner)
Fidelity	`mechanism_microport` — what the classes mean

This port is measured in the runtime matrix: the mechanism is preserved and measured under AgentDescent's runtimes; it is not a paper-benchmark reproduction.

The mechanism¶

AFlow searches the space of code-expressed agentic workflows. Its selection is not UCT: at the pinned revision it draws from the top-k scored workflows plus the always-included seed with mixed probability λ·uniform + (1−λ)·softmax(α·(s−s_max)), and keeps no visit counts (scripts/optimizer_utils/data_utils.py). Expansion asks an optimizer LLM to rewrite the selected workflow, with the parent's experience — its prior modifications and whether each helped — injected into the prompt.

Where each piece lives¶

Upstream mechanism	Where it lives here
Soft mixed selection over top-k + seed	`SoftMixed(SelectionPolicy)`, driven by the population aggregator over the archive of committed workflows
Per-father experience	a per-parent modification log injected into the expansion prompt
Workflow as code	two fixed model nodes (Solve → ReviewAndRevise) as `FieldSlots` keys
Convergence over 20 rounds	the candidate budget

Boundaries¶

Fixed two-node topology instead of code-level graph rewrites.
Paper hyper-parameters α=0.4, λ=0.2 (the pinned code itself ships 0.2/0.3).

Measured results — GSM8K¶

Three seeds, async_pipeline, 80 rollouts each, 8 workers, --staleness full, reflective merge on (this method's own declaration), deepseek-v4-flash at temperature 0.7. Recorded in bench/results/aflow-gsm8k.json.

seed	test quality	validation	accepted	calls	wall
0	0.609 → 0.984	0.609 → 0.969	1/80	2065	705 s
1	0.438 → 0.969	0.438 → 0.906	5/80	2071	600 s
2	0.484 → 0.953	0.469 → 0.969	4/80	2072	592 s

Mean final 0.969, mean gain +0.458, all three seeds moving.

64 items per split from GSM8K's own train and test splits. The baseline is 0.438–0.609 — the model's own accuracy — not the 0.000 the previous fixture installed by construction; see Self-Refine for why that fixture was replaced, and the caveat on PromptBreeder on one run per seed.

Two nodes cost two calls¶

~2070 calls per seed against Self-Refine's ~1050, for 0.969 against 0.943. AFlow's workflow is Solve then ReviewAndRevise, so every rollout is two model calls where Self-Refine's fused critique-and-refine is one.

On the old hand-written fixture the two were 0.896 and 0.979 — the gap ran the other way, and it was an artifact of that domain's one-bit output convention, which a second reviewing node could fix and a single instruction had to hold alongside the arithmetic. Against a benchmark with a real floor they land within 0.026 of each other, and what separates them is what each spends to get there. That reversal is the clearest thing the move to GSM8K bought.

Run it¶

python -m examples.aflow.aflow_workflow_search --dry-run

# one seed of the three above
python -m examples.aflow.aflow_workflow_search --yes --seed 0 \
    --provider openai --model deepseek-v4-flash \
    --async --async-ratio 1 --workers 8 --budget-rollouts 80 --staleness full \
    --temperature 0.7 --max-seconds 3600

--async-ratio 1 is what this row ran at: the flag was dropped before it reached the runtime, so the run took the runner's default whatever the command line said. It is passed explicitly here because the value matters and a default can move.

Flags: the MethodPolicy command line.

Offline tests: tests/test_aflow_upstream.py, tests/test_candidate_methods.py.