M2 — AFlow workflow topology search

Paper: Zhang, Xiang, Yu, Teng, Chen, Chen, Zhuge et al. — AFlow: Automating Agentic Workflow Generation (ICLR 2025 Oral)

Linear: AGE-7 · Crates: chatty-flow (runtime IR + interpreter), chatty-optimize (MCTS search) · Promotion: pending

Mechanism

Search over workflow topology before optimizing prompt text:

Nodes = LLM calls (model, prompt, temperature, format); edges = control flow
Operators: Generate, Format, Review & Revise, Ensemble, Test, Programmer, Custom
MCTS: each tree node is a complete workflow; soft-mixed selection (λ=0.2, α=0.4); LLM expansion with experience log; 5× validation evaluation; backpropagate diffs

Kill criterion: if IrRepr cannot express the search space by week 6, wait for Monty code mode or narrow the search — no third scripting language.

Chatty mapping

Paper concept	Chatty location	Status
Workflow representation	`WorkflowRepr` trait	Reserved (`WfNode`)
IR interpreter	`chatty-flow` → sub-agent / tool registry	To build
MCTS search	`chatty-optimize` (offline only)	To build
Composition substrate	`sub_agent_tool`, `invoke_agent_tool`, `list_agents_tool`	Ships
Monty code mode (future)	`sandbox/monty.rs`, `sandbox/monty_bridge.rs`	`MontySandbox` runs stdlib Python via a `python3` subprocess under rlimits (tried before Docker); `ToolBridge` is interface only — the pause/resume tool-call loop is not wired
Saved workflow	Future `FlowSettingsModel`	Gate: product integration

flowchart LR
  Search[chatty-optimize MCTS] --> IR[WorkflowRepr JSON]
  IR --> Interpreter[chatty-flow interpreter]
  Interpreter --> Tools[Existing tool registry]
  Tools --> Core[chatty-core agents]

Split is intentional: search never ships in the app binary; interpreter does.

Eval protocol

Stage A:

All 7 operators as composable Ops
Selection matches closed form for λ=0.2, α=0.4; blank template retains mass every round
Experience log stores parent diff, not just score
Discovered workflow serializes, reloads, re-executes to same score under warm cache

Stage B: HotpotQA / DROP / HumanEval / MBPP / GSM8K / MATH — target ordering (+5.7% over best manual design reported in paper; do not treat as binding until the human reviewer posts numbers).

Production landing

Mechanism	Likely promotion
MCTS search	Offline / CI only — reject from request path
`IrRepr` interpreter	Setting (`workflow_enabled`, `active_workflow_id`)
Winning workflow JSON	User-selected artifact via apply policy (AGE-45)

Reserved symbols

WorkflowRepr, WfNode, soft_mixed_select, SelectionStrategy — see RESERVED.md.

Depends on

M0 Trace, M1 ReAct
AGE-30 reflection gate (representation decision)

Chatty Documentation