Synthesis: APM Manifesto — core messages and critical read

Close-reading synthesis of Calvanese et al. 2026 (18 authors, arXiv:2603.18916v1, ~35 pp), produced 2026-04-21. Complements the existing source-page summary with a deeper analytic read: what the paper actually argues, how its parts fit together, and where a careful reader should push back.

Companion: syntheses/abpms-to-apm-evolution covers the paradigm shift from ABPMS (2023) to APM (2026). This page is about APM on its own terms.

1. The central thesis

Classical BPM/RPA assumes predefined workflows executed by passive software components. The emergence of LLM-based agents makes that paradigm obsolete for a growing class of processes. APM proposes that software and human agents are first-class functional entities that perceive, reason, and act within explicit normative frames. Autonomy and control are reconcilable — not opposed — through framing: deontic constraints (obligations, permissions, prohibitions) imposed on the agent's knowledge and goals, distinct from the procedural specifications (BPMN, DECLARE, RPA scripts) that BPM has used until now.

Explicit corollary, repeated in the Conclusions: LLMs alone are insufficient for intelligent BPM; symbolic frames remain essential.

2. The paradigm shift — what is made first-class

Dimension	Classical BPM / RPA	ABPMS (2023)	APM (2026)
Primary entity	Process (passive)	System augmented with AI	Agent (active, goal-driven)
Agency	Implicit — tools used	Implicit — embedded in system	Explicit — first-class
Control mechanism	Procedural (how to execute)	Procedural + AI heuristics	Normative frames (what is obligatory/prohibited)
Autonomy stance	Eliminated via automation	Constrained implicitly	Enabled within frames

Three abstractions are first-class in APM for the first time:

Normative frames — deontic specifications, distinct from operational frames (BPMN/DECLARE). See concepts/framed-autonomy.
Perceive–Reason–Act loop — canonical agent architecture with explicit perception, reasoning, action modules. See concepts/perceive-reason-act.
Conversational actionability — absent from ABPMS. Conversation is process-enactment, not a UI layer. See concepts/conversational-actionability.

3. The two-level architecture

Macro (APM System)

A socio-technical environment with a framing mechanism containing normative frames, shared process-level goals, and goal-alignment rules (hierarchies, coalitions, role assignment, segregation of duties). Agents interoperate via:

MCP (Model Context Protocol) — structured tool invocation
ACP (Agent Communication Protocol) — agent-to-agent messaging
RDF — semantic data exchange
KQML — knowledge-level communication

Agent categories: Human, Software (no deliberation), Embodied (robots), AI (deliberative via LLMs or other models). Tools include messaging, sensors, functions, services, actuators.

Micro (APM Agent)

A single agent runs a P-R-A loop with three conceptual modules:

Perception — sensors, percepts, belief-update; keeps internal world model current.
Reasoning (the core) — houses the four APM capabilities and holds:
Action — actuators, capabilities (skills/resources), ability to interact socially.

Grounded in BDI, FIPA, i*/Tropos, and foundational ontologies (DOLCE, GFO).

4. The four capabilities — why this ordering

The paper is explicit (p. 4) that the four capabilities are "purposefully ordered". Each is a prerequisite for the next; skip any earlier step and later steps become incoherent.

Capability 1 — Framed Autonomy

Operationally: the agent internalises deontic constraints into its knowledge and goals, then acts autonomously within them.

Distinctive move: BPMN and DECLARE, even when called "frames" in prior work, are operational — imperative about how to act. APM reframes frames as normative — deontic about what is required/permitted/forbidden.

Three blueprint scenarios:

Centralised — single decision-maker, frame on the process
Distributed, agent-level frames — each agent has its own frame
Distributed, process-level frames — hybrid constraints on process behaviour or parts

Why it comes first: without framing, there is nothing for the agent to be process-aware about.

Capability 2 — Explainability

Operationally: the agent articulates the rationale behind its decisions by specifying the explanans (what), the explanandum (when), and the XAI technique used.

Four desirable properties (§3.2):

Conforms to framing constraints and statements
Captures context richness
Reflects causal execution dependencies
Is interpretable to other agents (human or digital)

Distinctive vs. standard XAI: classical XAI explains one model's output. APM explanations must account for the frame (was a constraint respected or violated?), the context, and causal chains across a multi-agent process. In multi-agent settings, explanations must be legible to other agents, not just humans.

Why it comes second: autonomy without explanation is opacity. GDPR, EU AI Act, debugging, and trust all demand it.

Capability 3 — Conversational Actionability

Operationally: the agent interacts (via natural language, domain-specific languages, formal protocols like FIPA/KQML, token formats like TOON/MCP/ACP) and maps conversations to enactment.

Four enactment roles (§3.3):

Query — provide information on processes (model or execution level)
Recommend — suggest adaptations or future evolutions
Create — elicit process models from domain knowledge and data
Execute — trigger actions that move instances to new states

Distinctive move: classical BPM treats conversation as a UI on top of a process engine. APM makes conversation part of process enactment itself. Federated perspectives (multiple agents, potentially disagreeing) replace single-point-of-truth.

Why it comes third: framed and explainable agents still need to coordinate — receive instructions, report status, disagree and resolve.

Capability 4 — Self-Modification

Operationally: agents and coalitions adapt (short-term, instance-level) and evolve (long-term, model-level).

Dimension	Adaptation	Evolution
Scope	Single process instance	Across instances and contexts
Duration	Short-term, ephemeral	Long-term, persistent
Permanence	Does NOT alter process model	Alters process logic / model / policy
Knowledge type	Localised, instance-specific	Aggregated, systemic patterns
Mechanism	Individual or collaborative at runtime	Shared among multiple agents
Trigger	Reactive (deviation) or proactive (prediction)	Learned improvement opportunities

Feedback loop: successful adaptations become candidates for evolution; evolution reduces the need for future adaptation.

Why it comes fourth: self-modification presupposes all prior capabilities. Without framing, there is nothing to improve toward. Without explainability, improvements cannot be validated. Without conversational actionability, improvements cannot be communicated or coordinated.

5. Research agenda — 24 challenges thematically grouped

The paper enumerates 19 capability-specific + 5 cross-cutting challenges. Grouping by what each set is really asking:

F1–F4 (Framing): how to make frames real

F1 Pragmatic notion of an agent — the term "agent" has no shared definition in BPM practice post-LLM.
F2 Elicit and specify frames — requires a frame meta-model, specification language, and elicitation methods (from NL, rule-mining, domain experts). DECLARE and SIGNAL can be extended with deontic notions.
F3 Operationalise frames on real-world symbolic data — the short-to-mid-term prerequisite flagged by the authors. Real data is noisy, incomplete, inconsistent.
F4 Incentive and reinforcement mechanisms for frame-compliance — "frame compliance as a continuous, learnable property".

X1–X5 (Explainability): practical and trustworthy explanations

X1 Eliciting explanation preferences from stakeholders
X2 Leveraging and extending existing XAI techniques across the P-R-A loop
X3 Actionable explanations that preserve autonomy
X4 When to generate and how long to preserve explanations
X5 Causally sound explanations — spurious correlations are not valid explanations

A1–A5 (Actionability): bridging conversation and enactment

A1 Robust conversation with humans and other agents
A2 Exploiting process-management tools/services — twofold challenge of conversation-to-API mapping plus trust/specificity
A3 ABPMS modernising itself as an agent (wrapping legacy BPMS)
A4 Computing and using KPIs aggregated across multiple decision-making agents
A5 Balancing autonomy / delegation / control at the governance level

M1–M5 (Self-Modification): safe and effective continuous learning

M1 Governing self-modification — "learning to defer" (when to ask for help)
M2 Evaluating success/failure of modifications without human oversight
M3 Aligning concurrent evolution and adaptation across linked process executions
M4 Continuous learning and adaptation management — scaling to high-volume, long-running APM systems
M5 Modelling and measuring uncertainty (aleatoric vs. epistemic)

C1 Automated provisioning and onboarding of legacy BPM into APM — "agent-centric process mining"
C2 Security and privacy — prompt injection, unsafe learning, explanation-leakage
C3 Benchmarking and evaluation — benchmark contamination as methodological warning
C4 Liability and accountability — EU AI Act "responsibility gap"
C5 Engineering methods — no canonical APM method yet; build on AAII, Gaia, Tropos, Prometheus

The three the authors themselves flag as hardest

Reading §5 (Conclusions and Outlook):

Operationalising frames on real-world data (F3, C1, C2) — the prototype-to-production gap.
Self-modification at scale (M3, M4, M5) — concurrently evolving agents without divergence.
Bridging BPM and Multi-Agent Systems research traditions — the two communities have drifted; APM requires both.

6. Distinctions the paper insists on

Normative frame ≠ operational frame (deontic vs. procedural). "Do not delete a rejected assignment without notifying the student" vs. "retrieve ID → record in DB → send email." BPMN/DECLARE are operational.
Autonomy ≠ automation. Automation executes predefined tasks exactly; autonomy perceives, reasons, chooses within a frame.
Adaptation ≠ evolution. Instance-level ephemeral vs. model-level persistent.
Centralised ≠ distributed intelligence. Architectural choice, not a binary; frames can live on process, on agents, or both.
Process-aware ≠ task-centric. The agent understands it is part of a larger process with collective goals.

7. Positioning against prior work

PM Manifesto (2012) — successor-in-spirit. APM adopts the community-manifesto format and builds on process-mining foundations but shifts from retrospective analysis to prescriptive design and governance.
ABPMS Manifesto (2023) — direct predecessor. APM adopts the four ABPMS characteristics and reinterprets them as agent capabilities. ABPMS augments the system; APM governs the agent. See syntheses/abpms-to-apm-evolution.
Large Process Models (2024) — parallel vision. Concession: LLMs are valuable. Rejection: LLMs alone are insufficient; symbolic frames essential (p. 24-25).
"LLMs-with-tools" industry practice — tool-use alone does not create process-awareness, framing, or multi-agent governance.
Chapela-Campa & Dumas (2023) — four-level analytics pyramid (descriptive → diagnostic → predictive → prescriptive → augmented). Complementary, analytics-side lens; APM is agent-side.

8. What the paper explicitly does NOT claim

No reference implementation (p. 25)
No killer applications identified — the supplier-onboarding example (§2.1) is illustrative, not compelling
LLMs alone are not sufficient (repeated, p. 24–25)
No canonical engineering method — C5 calls for one
Regulatory frameworks are not settled — C4 treats EU AI Act liability as an evolving responsibility gap
Frame elicitation is not straightforward — F2 acknowledges all three elicitation approaches (domain experts, logs, NL) are incomplete

9. Critical read — strengths and thin spots

Strong

The normative/operational frame distinction is tight. This is the paper's real contribution. BPMN/DECLARE/RPA are reframed as operational, exposing a vacuum that normative frames fill. Concrete examples ground it.
The four-capability ordering is well-motivated. Each capability is a prerequisite for the next. This is a stack, not a list.
The macro/micro architecture is concrete enough. Figure 1 specifies what the framing mechanism contains, which protocols agents use, which modules comprise the P-R-A loop.
The 24 challenges are thematically organised, non-arbitrary. Each group asks a coherent question.
Positioning against prior work is explicit. The paper concedes what it adopts from 2012/2023/2024 predecessors and states clearly what it rejects.

Thin / aspirational

Frames remain partly abstract. F2 mentions DECLARE, SIGNAL, deontic logic, LLM-based generation — but does not commit to a formalism. A practitioner asks "in what language do I write a frame?" — the paper answers "it depends, and this is open."
Conversational actionability is underspecified. The four roles (Query/Recommend/Create/Execute) are useful, but how does an agent disambiguate a natural-language request into a role? A1 opens this, does not close it.
Self-modification is more aspirational than operational. The adaptation/evolution distinction is clear; the mechanics of learning are deferred to M1–M5. Reinforcement learning? Symbolic rule-learning? Hybrid? The paper stays agnostic.
LLMs' role in the reasoning module is unresolved. The paper cites refs [41, 42] characterising LLMs as "shallow disjunctive reasoners", yet uses LLM-based agents as its primary example. The tension is noted but not resolved.
No enterprise adoption path. C1 names "agent-oriented process mining" as the answer to migrating 20-year-old workflow systems; no concrete migration strategy is given.
Liability (C4) is flagged, not solved. A critical blocker for real-world deployment is left as an open research question.
No killer use case. Whether APM is urgently needed in practice, beyond being conceptually necessary, is less clear.

Tensions not fully resolved

Autonomy vs. control calibration. Too much framing kills autonomy; too little lets agents drift. F4 and A5 touch on this; neither resolves it.
Centralised vs. distributed intelligence. Both presented as viable (§3.1); tradeoffs not deeply discussed. Distributed constraint optimization is mentioned for M3 but not evaluated.
LLM capability ceiling. If LLMs are weak reasoners, what is the agent's reasoning architecture? The answer implied is "frames + LLMs", but the integration depth is unclear.

10. Bottom line

The APM Manifesto succeeds as a manifesto — it articulates a vision, calls a community to action, and outlines a substantive research agenda. It does not succeed as a practitioner guide, and it does not claim to. The paper's strongest move is the normative-vs-operational frame distinction, which recasts the BPM problem and explains why BPMN/RPA are insufficient for autonomous agent governance. Its weakest area is operationalisation — the gap between "agents should be framed and process-aware" and "here is how to implement that in your organisation" remains large.

The critical test of the manifesto is whether the 24 challenges get taken up and delivered over the next few years. Papers like sources/2025-calvanese-autonomy-business-process-execution (operationalising frames), sources/2025-fournier-agentic-ai-process-observability (operationalising observability), and sources/2025-elyasaf-self-modifying-abps (operationalising self-modification) are the early returns on that agenda.

concepts/agentic-bpm — hub concept for the APM paradigm
concepts/framed-autonomy, concepts/explainability-apm, concepts/conversational-actionability, concepts/self-modification, concepts/perceive-reason-act — capability-specific concept pages
syntheses/abpms-to-apm-evolution — what changed from 2023 to 2026
syntheses/apm-business-themes — APM agenda mapped to 12 business themes
syntheses/llm-bpm-reading-list — broader reading list for LLMs in BPM