Opus Workflow Intention Framework

• Opus Workflow Intention Framework is a structured approach that defines workflow objectives by aligning input, process, and output signals.
• It integrates multimodal encoding and attention layers to extract signals, decode intentions, and drive robust workflow generation and optimization.
• The framework enables separation of mixed-intent queries into distinct workflows, improving generation metrics and handling complex business processes.

The Opus Workflow Intention Framework denotes a line of work in which workflows are generated, interpreted, or optimized through an explicit representation of intention rather than from raw prompts alone. In its formal multimodal version, Workflow Intention is defined as “the alignment of Input, Process and Output elements defining a Workflow's transformation objective interpreted from Workflow Signal inside Business Artefacts” (Kingston et al., 25 Feb 2025). In its prompt-oriented formulation, the same idea appears as an intermediate Intention Capture layer between user queries and Workflow Generation, built from Workflow Signals and Workflow Intention objects (Fagnoni et al., 15 Jul 2025). A later Opus system for complex Business Process Outsourcing use cases reformulates intention as the structured alignment of Client Input, Process Context, and Client Output, and uses that object to drive workflow generation and optimization (Fagnoni et al., 2024).

1. Definitions, scope, and variant formulations

Across the Opus literature, “intention” is not a generic synonym for user preference. It is a structured description of a workflow’s transformation objective. The common thread is that a workflow should be generated from an explicit intermediate representation of what enters the process, what transformation is required, and what result is expected. In the formal Workflow Intention paper, that representation is the triple of Input, Process, and Output (Kingston et al., 25 Feb 2025). In the prompt-intention paper, it is operationalized as Workflow Signals extracted from user queries and then grouped into Workflow Intention objects before any workflow is produced (Fagnoni et al., 15 Jul 2025). In the large-work-model paper, the formulation is widened to Client Input, Process Context, and Client Output, emphasizing operational context and downstream graph optimization (Fagnoni et al., 2024).

Paper	Intention formulation	Primary role
"Opus: A Workflow Intention Framework for Complex Workflow Generation" (Kingston et al., 25 Feb 2025)	Alignment of Input, Process and Output elements interpreted from Workflow Signal inside Business Artefacts	Formal representation and multimodal generative architecture
"Opus: A Prompt Intention Framework for Complex Workflow Generation" (Fagnoni et al., 15 Jul 2025)	Workflow Signals from user queries, interpreted into structured Workflow Intention objects	Intermediate prompt-processing layer for LLM workflow generation
"Opus: A Large Work Model for Complex Workflow Generation" (Fagnoni et al., 2024)	Alignment of Client Input, Process Context, and Client Output	End-to-end generation and optimization with WKG and LWM

This suggests a family of closely related formulations rather than a single frozen specification. The invariant idea is that workflow generation should proceed through an explicit intention object, not through direct prompt-to-workflow mapping alone. A second invariant is separation: if a query contains multiple distinct transformation objectives, those objectives should be separated into multiple intentions and, correspondingly, multiple workflows (Fagnoni et al., 15 Jul 2025).

2. Formal representation of Workflow Signals and Workflow Intentions

The most explicit mathematical treatment appears in the formal Workflow Intention framework. It defines a Workflow Signal space $\text{S}$ as a subspace of a finite-dimensional real Hilbert space $\text{X}$, together with three subspaces for Input, Process, and Output such that

$$\text{S} = \text{I} \oplus \text{P} \oplus \text{O}.$$

Any signal $s \in \text{S}$ decomposes uniquely as

$$s = i + p + o, \quad i \in \text{I},\; p \in \text{P},\; o \in \text{O}.$$

The corresponding Workflow Intention space is the product space

$$\mathcal{G} = \text{I} \times \text{P} \times \text{O},$$

and a single Workflow Intention is a point

$$\gamma = (i_\gamma, p_\gamma, o_\gamma) \in \mathcal{G}.$$

This formalism distinguishes fragmentary evidence from coherent transformation objectives. A Workflow Signal is a local informational cue about input, process, or output; a Workflow Intention is a complete triple aligning those components into a single transformation objective (Kingston et al., 25 Feb 2025).

The same paper introduces generative families $\text{I}_g$, $\text{P}_g$, and $\text{O}_g$, which function as canonical sets of input, process, and output elements. These families support classification heads that map neural representations into interpretable counts over canonical IPO elements. The overall mapping from context and signals to intentions is written as

$\text{X}$0

so that encoded multimodal context and extracted signals yield a set of intentions rather than a single compulsory output. This is important because the framework explicitly allows mixed or multiple intentions in the same artefact collection (Kingston et al., 25 Feb 2025).

A later prompt-oriented formulation simplifies the same idea for LLM prompting. There, a Workflow Signal is a triple $\text{X}$1 of lists of strings, and each Workflow Intention is likewise a triple of Input, Process, and Output lists. For sample $\text{X}$2, a reference Intention Set is written as

$\text{X}$3

where $\text{X}$4 is the Mixed Intention Level, i.e., the number of distinct transformation objectives embedded in the query (Fagnoni et al., 15 Jul 2025). The simplification from Hilbert-space objects to list-structured prompt objects is pragmatic rather than contradictory: it preserves the IPO alignment while making the representation directly usable by instruction-tuned LLMs.

3. Multimodal intention resolution architecture

The formal Opus architecture is organized into four stages: Modality-Specific Encoding, Intra-Modality Attention, Inter-Modality Fusion Attention, and Intention Decoding (Kingston et al., 25 Feb 2025). It supports at least three modalities in the paper: Text, Image, and Document. Each modality is processed by its own tokenizer, embedding layer, encoder, and unifying projection into a shared latent dimension $\text{X}$5. From each modality-specific representative vector, three projection heads produce vector-valued Input, Process, and Output signals.

For text artefacts, the framework uses a RoBERTa-large tokenizer and encoder with a special $\text{X}$6 token, followed by linear projections to shared-dimensional IPO vectors. For image artefacts, it uses an InternViT-style tiling and encoding pipeline, followed by projection into the shared space. For documents, it combines text, image, and spatial information page-wise, representing text boxes and image regions and encoding them into a unified document representation before projecting to IPO signals (Kingston et al., 25 Feb 2025).

After per-artefact encoding, Intra-Modality Attention aggregates multiple artefacts of the same modality. The model concatenates encoded artefact sequences for a given modality and applies a modality-specific transformer to produce aggregated Input, Process, and Output signals for the entire modality. Inter-Modality Fusion then concatenates these modality-level representations and feeds them to a fusion encoder, yielding a multimodal latent context $\text{X}$7 (Kingston et al., 25 Feb 2025).

Intention Decoding is autoregressive. At each step, the decoder produces an intermediate vector $\text{X}$8, maps it to IPO vectors, and evaluates two stopping mechanisms: a stopping head and a redundancy criterion based on average cosine similarity across previously generated intentions. The redundancy rule is designed to enforce non-redundancy among intentions, consistent with the earlier formal requirement that distinct intentions must differ in at least one component (Kingston et al., 25 Feb 2025).

Training is divided into two phases. Phase 1 trains the modality-specific and intra-modality encoders for signal extraction. Phase 2 freezes those components and trains the fusion encoder, decoder, stopping head, and intention-level components. The principal losses are a bounded signal loss,

$\text{X}$9

and an overall intention loss,

$$\text{S} = \text{I} \oplus \text{P} \oplus \text{O}.$$0

The sequence term combines coverage, overlength, and underlength penalties for generated intention sets, while the contrastive term discourages near-duplicate intentions (Kingston et al., 25 Feb 2025).

4. Prompt intention capture and mixed-intention workflow generation

The prompt-based Opus formulation recasts the same problem as an intermediate layer between user queries and workflow generation. Instead of directly mapping a query to a workflow, the system performs two LLM-mediated operations: Signal Extraction and Intention Generation (Fagnoni et al., 15 Jul 2025). For a mixed user query $$\text{S} = \text{I} \oplus \text{P} \oplus \text{O}.$$1, Signal Extraction produces

$$\text{S} = \text{I} \oplus \text{P} \oplus \text{O}.$$2

where $$\text{S} = \text{I} \oplus \text{P} \oplus \text{O}.$$3, $$\text{S} = \text{I} \oplus \text{P} \oplus \text{O}.$$4, and $$\text{S} = \text{I} \oplus \text{P} \oplus \text{O}.$$5 are extracted lists of input, process, and output elements. Intention Generation then produces an Intention Set

$$\text{S} = \text{I} \oplus \text{P} \oplus \text{O}.$$6

Workflow generation is then performed either directly from the query,

$$\text{S} = \text{I} \oplus \text{P} \oplus \text{O}.$$7

or from the predicted intentions,

$$\text{S} = \text{I} \oplus \text{P} \oplus \text{O}.$$8

The paper formulates two design principles. First, incomplete intentions should be flagged rather than silently completed during workflow generation. Second, mixed intentions should be separated so that the model generates one workflow per intention rather than one entangled DAG for several unrelated transformations (Fagnoni et al., 15 Jul 2025). The benchmark is built precisely to test that claim: it contains 1,000 multi-intent query–workflow pairs with Mixed Intention Level $$\text{S} = \text{I} \oplus \text{P} \oplus \text{O}.$$9, i.e., between one and ten distinct transformation objectives per query.

Evaluation uses signal loss, intention loss, standard semantic workflow similarity metrics, and LLM-as-a-Judge scoring. The empirical result is that direct workflow generation degrades sharply as query complexity rises, whereas intention-guided generation remains comparatively stable. For Claude 3.7 Sonnet, ROUGE-1 without intention drops from $s \in \text{S}$0 at $s \in \text{S}$1 to $s \in \text{S}$2 at $s \in \text{S}$3, while with intention it remains around $s \in \text{S}$4 at $s \in \text{S}$5 and around $s \in \text{S}$6 at $s \in \text{S}$7. Over the same range, cosine similarity without intention drops from about $s \in \text{S}$8 to $s \in \text{S}$9, whereas with intention it remains around $$s = i + p + o, \quad i \in \text{I},\; p \in \text{P},\; o \in \text{O}.$$0 at $$s = i + p + o, \quad i \in \text{I},\; p \in \text{P},\; o \in \text{O}.$$1 (Fagnoni et al., 15 Jul 2025).

The framework also defines set-level workflow similarity by matching generated workflows to reference workflows through injections that maximize total similarity. That construction is necessary because mixed-intention queries may yield workflow sets rather than a single graph. This makes the evaluation consistent with the underlying one-intention-per-workflow principle (Fagnoni et al., 15 Jul 2025).

5. From intention to workflow graphs, optimization, and runtime selection

A later Opus system moves from intention capture to end-to-end workflow synthesis and optimization for complex BPO use cases. In that formulation, Intention is “the structured alignment of the Client’s Input (I), potential Process Context, and the expected or actual Client’s Output (O)” (Fagnoni et al., 2024). The architecture has four layers: Intention Encoding, Workflow Graph Generation, Workflow Optimization, and Workflow Execution. Two global components support these layers: a Work Knowledge Graph (WKG) and a Large Work Model (LWM).

The intention machinery is multimodal. For each modality $$s = i + p + o, \quad i \in \text{I},\; p \in \text{P},\; o \in \text{O}.$$2, preprocessed inputs are encoded as

$$s = i + p + o, \quad i \in \text{I},\; p \in \text{P},\; o \in \text{O}.$$3

and then fused by an Intention Encoder into a joint vector

$$s = i + p + o, \quad i \in \text{I},\; p \in \text{P},\; o \in \text{O}.$$4

An Intention Decoder transforms $$s = i + p + o, \quad i \in \text{I},\; p \in \text{P},\; o \in \text{O}.$$5 back into a structured textual description of the Client Input, Client Output, and required process. Intention Routing compares $$s = i + p + o, \quad i \in \text{I},\; p \in \text{P},\; o \in \text{O}.$$6 with embedded WKG nodes via cosine similarity and selects a relevant node set $$s = i + p + o, \quad i \in \text{I},\; p \in \text{P},\; o \in \text{O}.$$7 (Fagnoni et al., 2024).

Workflow generation then conditions the LWM on both the decoded intention text and textualized WKG subgraphs. The generated workflows are DAGs whose nodes are Tasks; each Task is defined as a sequence of executable Instructions, which can include tools, code, LLM agent invocations, and human expert reviews. Candidate DAGs are merged into a Workflow Graph $$s = i + p + o, \quad i \in \text{I},\; p \in \text{P},\; o \in \text{O}.$$8, after which Opus solves a path optimization problem

$$s = i + p + o, \quad i \in \text{I},\; p \in \text{P},\; o \in \text{O}.$$9

This separates semantic alignment from cost-aware selection: generation constrains the search space to intention-compatible workflows, and optimization selects among those candidates according to compute, latency, and model-usage costs (Fagnoni et al., 2024).

The same broad design appears in a later evaluation framework that quantifies workflow quality and efficiency by combining an expected-value reward with normative penalties for Cohesion, Coupling, Observability, and Information Hygiene. There, a workflow is modeled as a probabilistic DAG $$\mathcal{G} = \text{I} \times \text{P} \times \text{O},$$0, and reward is defined as expected output gain minus cost, while normative penalties provide structural and signal-quality regularization (Seroul et al., 6 Nov 2025). This suggests that an Opus intention layer is compatible not only with generation but also with downstream ranking and optimization criteria.

6. Related research, scope boundaries, and limitations

The Opus formulation sits within a broader research landscape in which workflow behavior is made intention- or goal-sensitive through explicit representations, metrics, or postconditions. A collaboration-centered architecture for workflow-driven team science proposed PPoDS and SmartFlows as a way to capture intentions early as metrics and test cases, then use them for runtime monitoring, prediction, and optimization (Altintas et al., 2019). A linked-data workflow framework represented each atomic activity by a postcondition expressed as a SPARQL ASK query, effectively treating workflow nodes as goal conditions over world state (Käfer et al., 2018). An interactive writing system, IntentFlow, distinguished stable goals from fluid intents and represented intents as editable objects with explicit links to output segments, showing that persistent intent representations improve alignment and reduce corrective prompting (Kim et al., 29 Jul 2025). A later workflow-construction framework, FlowMind, argued that execution and workflow construction should be decoupled: first solve the task with tools, then reconstruct the workflow from execution traces, because simultaneous execution and workflow construction create interference and reduce workflow fidelity (Liu et al., 12 Feb 2026).

Several limitations recur across the Opus papers. The prompt-intention benchmark is synthetic, even though it is large and systematically varied (Fagnoni et al., 15 Jul 2025). The formal multimodal intention paper is architectural and mathematical rather than empirically benchmark-driven; it specifies losses and training stages but does not report a comparable experimental section (Kingston et al., 25 Feb 2025). The large-work-model paper explicitly notes that workflow execution is outside its scope and that mis-inference is possible when Client Output or Process Context is missing (Fagnoni et al., 2024). More generally, the literature suggests that incomplete intentions, ambiguous artefacts, and weak coverage in the knowledge graph remain principal failure modes.

A common misconception is terminological. A separate paper titled "OPUS: an interoperable job control system based on VO standards" describes a UWS-based asynchronous job control system and explicitly states that it is not a full-blown workflow engine (Servillat et al., 2021). That system is relevant to interoperability and provenance, but it is distinct from the Workflow Intention line of work.

Taken together, the Opus Workflow Intention Framework can be understood as a program of making workflow objectives explicit, structured, and machine-operable before generation. In its strongest form, that program has three layers: extract signals from artefacts or prompts, resolve those signals into one or more aligned intention objects, and only then generate, optimize, or evaluate workflows against those objects.