Synthesis: From ABPMS to APM — what changed

A side-by-side comparison of the AI-Augmented BPMS Manifesto (Dumas et al. 2023, ACM TMIS) and its agent-centric successor, the Agentic Process Management Manifesto (Calvanese et al. 2026, arXiv). Authored 2026-04-13.

1. Core paradigm shift: system-augmentation → agent-centricity

	ABPMS (2023)	APM (2026)
Subject of augmentation	The system (BPMS) is augmented with AI	The agent is the primary functional entity
Position	"AI-augmented information system"	"Agentic socio-technical system"
Agency	Implicit — agents used, not first-class	Explicit — first-class citizens (human / software / embodied / AI)
Driving question	"How do we augment a BPMS with AI?"	"How do we govern autonomous agents within a process frame?"

The APM Manifesto states this explicitly: ABPMS "leaves the agentic perspective implicit". APM makes it explicit. This is the central shift.

2. Characteristics → capabilities (3 + 1 mapping)

ABPMS's four characteristics of the system become APM's four capabilities of the agent — plus one new dimension:

ABPMS characteristic	→	APM capability
Adaptable	→	concepts/self-modification (adaptation aspect)
Proactive	→	concepts/self-modification (evolution aspect) + concepts/perceive-reason-act
Explainable	→	concepts/explainability-apm (preserved, elevated to first-class)
Context-sensitive	→	concepts/framed-autonomy + concepts/process-awareness
(no counterpart)	+	concepts/conversational-actionability ← genuinely new

Conversational actionability is the new dimension absent from ABPMS — agents must converse (with humans and other agents) and link conversation to enactment (Query / Recommend / Create / Execute).

APM also imposes an explicit ordering of capabilities (framing → explainability → actionability → self-modification) that ABPMS does not — because ABPMS describes the system, while APM constructs the agent layer by layer.

3. "Frame" — terminological re-framing

	ABPMS	APM
Frame =	Operational specification (≈ BPMN/DECLARE)	Normative specification (deontic)
Content	How the agent should act	What the agent must / must not do
Example	"Send email → update DB → archive"	"Never delete a rejected submission without notifying the student"

APM explicitly distinguishes operational frame vs normative frame — and identifies ABPMS's frames as operational in this terminology. A sharpening of the theoretical foundation.

4. New explicit architecture

ABPMS did not specify an explicit agent architecture. APM adds:

Macro level (APM System): the "world", framing mechanism, goal alignment, protocols (MCP, ACP, RDF, KQML).
Micro level (APM Agent): Perceive–Reason–Act loop with three modules (perception, reasoning, action).
Foundational ontologies & frameworks: DOLCE, GFO, BDI, FIPA, i*/Tropos.
Agent categories: human, software, embodied, AI.

5. Self-modification: new explicit distinction

APM introduces the clean split adaptation vs evolution absent from ABPMS:

Adaptation — short-term, instance-specific, non-persistent.
Evolution — long-term, shared across instances, alters the process model itself.

This is APM's mechanism to operationalise what ABPMS called "adaptable".

6. Research agenda — tightened and broadened

	ABPMS	APM
Challenges	"A set of challenges" (not enumerated by name in abstract)	19 named (F1–F4, X1–X5, A1–A5, M1–M5) + 5 cross-cutting (C1–C5)
New themes in APM	—	Prompt-injection security (C2), benchmark contamination (C3), EU AI Act / liability (C4), engineering methods (C5), agent-oriented process mining (C1)
Frameworks built on	Trustworthy AI (general)	LLM agents, MAS, MCP/ACP, deontic logic, normative MAS, "learning to defer"

7. Author continuity — same community, next generation

Of the 14 ABPMS authors, 8 recur in the APM Manifesto: Dumas, Fournier, Limonad, Marrella, Montali, Calvanese, De Giacomo, La Rosa.

The 10 new APM authors come predominantly from the agent / MAS tradition (Sardiña, Kampik, Polyvyanyy, Tax) — mirroring the conceptual shift from system to agent perspective. APM is not a competing paradigm — it is the next-generation continuation of the same intellectual collective.

8. What did NOT change

Trustworthy AI as ethical anchor — preserved.
Explainability as first-class — preserved (in fact reinforced in APM with 4 desirable properties).
Process-aware information system as substrate — preserved; APM builds on PAIS, not in place of it.
BPM lifecycle as reference frame — preserved; APM re-interprets but does not reject it.

9. One-sentence summary

ABPMS asked "how do we add AI to a BPMS?"; APM asks "how do we govern autonomous agents acting under normative frames?" — and answers it by elevating the agent to first-class status, reframing ABPMS's "frames" as operational (vs the new normative APM frames), reordering ABPMS's four characteristics into four ordered agent capabilities, and adding conversational actionability as a fifth dimension that ABPMS did not foreground.

10. Practical use of this synthesis

For new readers: read ABPMS first (system view), then APM (agent view). The terminology only fully clicks after seeing both.
For comparing future BPM/AI vision papers: use the ABPMS quartet and the APM capability stack as anchor frames.
For research scoping: APM's named challenges (F1–F4, X1–X5, A1–A5, M1–M5, C1–C5) are the most concrete current research agenda — most ABPMS-era questions are subsumed.

Connections

Sources: sources/2023-dumas-ai-augmented-bpms · sources/2026-calvanese-agentic-bpm-manifesto Concepts: concepts/agentic-bpm · concepts/framed-autonomy · concepts/process-awareness · concepts/explainability-apm · concepts/conversational-actionability · concepts/self-modification · concepts/perceive-reason-act