SDOF: Taming the Alignment Tax in Multi-Agent Orchestration with State-Constrained Dispatch¶
๐ Published (v1): 2026-04-20 12:51 UTC ยท Source: Arxiv ยท link
Why this paper was selected
State-constrained dispatch fixes alignment tax in multi-agent orchestration frameworks
Ask a follow-up
Open an assistant pre-loaded with this paper's context.
๐ฌ Ask ChatGPTโฆ Ask Claude
TL;DR¶
SDOF wraps multi-agent LLM pipelines in a two-layer harness: a GRPO-trained 7B Intent Router for intent recognition, and a StateAwareDispatcher that enforces a GoalStage finite-state machine plus SkillRegistry preconditions before any API call executes. Evaluated on a live enterprise HR platform (Beisen iTalent, 1,671 real API calls), SDOF blocks 100% of injected illegal operations while achieving 86.5% task completion โ +11.9% over an unconstrained baseline.
Problem¶
Existing orchestration stacks (LangChain, LangGraph, CrewAI, AutoGen) expose only transition graphs; they do not enforce whether the current workflow stage permits a requested action. A SOURCING-phase agent can invoke an interview-scheduling API if a graph edge exists, violating business-process ordering rules and causing compliance or legal failures in regulated industries. No prior framework simultaneously provides intent-stage binding orthogonal to transition graphs, precondition validation at skill level, and auditable execution traces validated against live production APIs.
Method¶
SDOF models multi-agent execution as a constrained state machine \(G = (S, s_0, T, \delta, I, \Lambda)\), where \(\Lambda : I \to 2^S\) is the intent-stage binding โ each intent is legal only in the stages \(\Lambda(i) \subseteq S\), orthogonal to the transition relation \(T\).
Layer 1 โ Online-RLHF Intent Router: A 7B model fine-tuned with GRPO (veRL) to output structured intent classifications; falls back from deterministic string-match (0.12 ms, 97.5% STA) to LLM routing for ambiguous inputs.
Layer 2 โ StateAwareDispatcher + SkillRegistry (Algorithm 1):
1. Router identifies intent \(i\); dispatcher calls R.select_skill(intent, s) โ stage-filtered against \(\Lambda\).
2. Preconditions \(\Pi_{pre}\) are checked: \(\forall \pi \in \Pi_{pre}(sk),\ \pi(C) = \top\).
3. On pass, skill executes; postconditions applied; stage transitions via G.advance_stage only if the target is reachable.
4. Every step emits a replayable ProcessEvent trace to the GoalManager (PostgreSQL).
SkillRegistry uses a three-level risk taxonomy (L0 atomic, L1 composite, L2 policy) with progressive disclosure: low-context L0 manifests exposed at routing time; richer L1/L2 descriptions loaded only after binding.
The paper also formalizes the alignment tax: \(\Delta_{tax} = P(\Phi \mid x, \Pi_{base}) - P(\Phi \mid x, z, \Pi_{think})\), the structural-validity drop induced by chain-of-thought tokens in reasoning-mode LLMs, which SDOF mitigates by enforcing hard dispatcher checks independent of model output quality.
Key Contributions¶
- Intent-stage binding \(\Lambda\) formalization as a constraint layer orthogonal to transition graphs โ enforces which intents are legally executable at each workflow stage.
- SDOF framework: GoalStage FSM + SkillRegistry (with \(\Pi_{pre}\)/\(\Pi_{post}\)) + StateAwareDispatcher packaged into a reusable orchestration layer.
- GRPO-aligned 7B Intent Router outperforming GPT-4o zero-shot (80.9% vs 48.9% joint accuracy) on an FSM-constrained adversarial routing benchmark.
- Production evaluation on 185 HR scenarios with 1,671 live Beisen iTalent API calls; cross-domain validation on 960 SGD-derived dialogues across 8 service domains.
- Per-step trace grading: 882 dispatcher records with intent, stage, skill, precondition result, and outcome โ enabling fine-grained error attribution beyond pass/fail scenario metrics.
Results¶
- TCR: SDOF 86.5% (95% CI 80.8โ90.7) vs LangGraph 84.9%, LangGraph+Pre 88.1%, Vanilla 74.6%.
- Illegal operations blocked: 22/22 (100%) in injected-illegal HR subset; LangGraph blocks 0/22.
- CVR: SDOF 2.5% vs 19.8% without StageCheck (+696% relative increase on ablation); ablation without precondition checks yields CVR 2.2% but drops TCR (preconditions guard a smaller, semantically unsafe subset).
- Blocking audit: Precision 100%, recall 88%, expert agreement ฮบ = 0.94.
- Intent Router: 80.9% joint accuracy vs GPT-4o zero-shot 48.9% on the 47-example FSM routing eval.
- Cross-domain (SGD): 201 stage-order conflicts detected across 960 dialogues, 41 in the normal (non-adversarial) split; all 160 injected illegal ops blocked.
- Dispatch latency: 57.4 ms (SDOF, includes live API); legality check overhead < 1 ms; LangGraph baseline 1.1 ms (no live API).
- Trace composition: 860/882 steps SUCCESS (97.5%), 16 ILLEGAL TRANSITION (1.8%), 6 PRECONDITION FAIL (0.7%).
Limitations¶
- FSM stage definitions are manually authored; no automated discovery from process logs (process-mining direction identified but not pursued).
- Baselines AutoGen and MetaGPT excluded from like-for-like comparison due to instrumentation incompatibility with the legality-trace protocol.
- Ablation isolates dispatcher governance; the standalone contribution of the intent router and its interaction with dispatcher checks is left to future experiments.
- Evaluated on a single enterprise tenant (Beisen 430008) and one HR domain for the primary benchmark; cross-domain SGD uses synthetic FSM mappings rather than production APIs.
- Extended multi-seed training comparisons and deeper-workflow evaluations deferred to a subsequent paper update (arXiv v1 scope caveat).
- LangGraph+Pre achieves comparable CVR (2.8%) with manually configured preconditions, challenging SDOF's automation claim.
Relevance to Harnesses / Meta-Harnesses¶
SDOF self-describes its architecture as a "harness control architecture" and explicitly positions itself against "harness-style runtimes" that handle task continuity, memory injection, and scheduling โ arguing SDOF occupies a narrower niche focused on workflow legality rather than operational breadth. The framework is a concrete instantiation of the meta-harness pattern: a deterministic orchestration wrapper that constrains, audits, and governs an underlying generative model rather than extending its capabilities. For researchers tracking harnesses, SDOF demonstrates how FSM-based constraint layers can be bolted onto existing orchestration graphs (LangGraph-style) without modifying individual agents โ a composable governance shell. Its four-role memory decomposition (working, procedural, reference, audit) and per-step trace grading also advance the operational primitives that meta-harness designs must account for.