RFC: Declarative (JSON/YAML) definition of Action-Driven Networks with per-turn structured-output enforcement

[elCaptnCode]

We build a framework (Mozaiks, OSS) that treats AG2 as its agentic execution backbone. Our workflows are declarative contracts: agents, tools, transition rules, and structured-output schemas are declared in strict YAML, validated against typed models, and then compiled at runtime into AG2-native objects (TransitionGraph, WorkflowAdapter, ContextEquals, ToolCalled, etc.) through a thin adapter. Determinism is the core requirement — routing decisions and artifact shapes must be reproducible from the declared contract, with the LLM's freedom confined to producing schema-validated structured outputs.

The Action-Driven Network design (blog post) fits this well conceptually — stateless adapters, state as a pure fold of the WAL. But today we maintain several adapter-layer workarounds to keep the declarative/deterministic contract, and we'd like to know which of these AG2 intends to own natively:

1. Declarative network definition. Is a canonical serialized schema (JSON/YAML) planned for defining a Workflow channel — agents, TransitionGraph rules, adapter config — so a network can be loaded, validated, and diffed as data rather than constructed only in Python? Today we compile our own YAML into the Python objects; a first-class serialized form (or a stable "compile target" schema) would let external frameworks target the Action-Driven Network without private adapters.

2. Per-agent response_schema on Workflow channels. Agent.ask() supports response_schema, but network turns don't appear to. We currently validate structured outputs after reading the WAL and treat our validation as the artifact authority. Is provider-level schema enforcement per agent turn on Workflow channels planned? (We'd still keep our own validation as the contract authority — we want model-side pressure, not replacement.)

3. Composable, data-declarable transition conditions. Our contracts declare source-scoped deterministic routes (source_agent + context equality / tool-called / small context-variable expressions). AG2's built-in conditions don't compose (FromSpeaker AND ContextEquals) and there's no expression evaluator, so we register custom conditions. Is native condition composition and/or a serializable condition grammar planned, so routing rules can live entirely in the declared file?

4. A public round-end packet hook. Our tools mutate context variables that routing must see before WorkflowAdapter.fold(...) selects the next speaker; we currently wrap the default notify handler to merge those updates into the round-end packet, which is our most fragile divergence. Is a supported hook or native context-update path planned?

5. Typed one-shot Consulting primitive. Many of our control-plane checkpoints are exactly the Consulting shape (one question, one schema-validated reply, hard close). Is a typed one-shot Consulting API with response_schema planned, so these don't need to go through Agent.ask() outside a channel?

The overall question: does AG2 see "the network as a validated data artifact" — declarative definition, typed conditions, per-turn schema enforcement — as in scope for the Action-Driven Network roadmap, or should frameworks like ours plan to permanently own the compile-from-declaration layer and keep only the runtime execution on AG2? Either answer helps us draw the boundary correctly; we'd rather consume native primitives than maintain parallel ones.

Happy to share our adapter code and YAML contract shapes if useful.