---
title: "AI-Agent Redesign Heuristics (2026) — re-reading Reijers & Liman Mansar (2005) for the agentic era"
type: synthesis
tags: [bpr, redesign, heuristics, devils-quadrangle, agentic-bpm, apm, ai-agents, 2026]
sources:
  - "[[sources/2005-reijers-limanmansar-best-practices-bpr]]"
  - "[[sources/2018-dumas-fundamentals-of-bpm]]"
  - "[[sources/2026-calvanese-agentic-bpm-manifesto]]"
  - "[[sources/2026-dumas-agentic-bpms-pyramid]]"
  - "[[sources/2025-calvanese-autonomy-business-process-execution]]"
  - "[[sources/2025-elyasaf-self-modifying-abps]]"
  - "[[sources/2025-fournier-agentic-ai-process-observability]]"
  - "[[sources/2025-vu-practitioner-perspectives-agent-governance]]"
  - "[[sources/2022-kubrak-prescriptive-ppm-slr]]"
  - "[[sources/2024-kampik-large-process-models]]"
created: 2026-05-12
updated: 2026-05-13
---

# AI-Agent Redesign Heuristics (2026) — re-reading the 2005 catalogue for the agentic era

A synthesis that takes up the open call in [[sources/2026-dumas-agentic-bpms-pyramid|Dumas, Milani & Chapela-Campa (2026, p. 45)]] for an **"agentic extension of the classical Reijers & Liman-Mansar redesign heuristics."** It reinterprets the 29 heuristics of [[sources/2005-reijers-limanmansar-best-practices-bpr|Reijers & Liman Mansar (2005)]] in light of (i) software agents executing the [[concepts/perceive-reason-act|Perceive–Reason–Act]] loop within [[concepts/framed-autonomy|normative frames]] ([[sources/2026-calvanese-agentic-bpm-manifesto|APM Manifesto, Calvanese et al. 2026]]) and (ii) generative/agentic AI as a process performer ([[sources/2026-dumas-agentic-bpms-pyramid|A-BPMS keynote, Dumas et al. 2026]]). Companion HTML map: [[assets/ai-agent-heuristics-devils-quadrangle.html]].

## Why redo this work in 2026?

Three forces unsettle the 2005 catalogue's trade-off profiles:

1. **The marginal cost of cognition collapses.** Heuristics like *parallelism*, *triage*, *task automation*, *buffering*, *contact reduction* were costed against scarce human attention. Once an LLM-mediated agent supplies cognition on demand, several of those costs fall close to zero — and others (latency, hallucination, alignment) appear where they did not exist before.
2. **The actor stops being passive.** [[sources/2026-calvanese-agentic-bpm-manifesto|Calvanese et al. (2026)]] insist that APM treats agents as primary functional entities with goals and the autonomy to perceive, reason and act inside an explicit frame. Heuristics that prescribe *which resource* performs a task (specialist/generalist, case manager, empower, order assignment) acquire a new degree of freedom: the resource itself can negotiate the frame.
3. **Design-time and run-time fuse.** [[sources/2025-elyasaf-self-modifying-abps|Elyasaf et al. (2025)]] split self-modification into **adaptation** (instance-level, ephemeral) and **evolution** (cross-instance, persistent). A "heuristic" therefore has two enactment surfaces: it can be applied at redesign time by a human modeller, *or* it can be applied at runtime by a self-modifying agent. The 2005 catalogue assumed only the former; the prescriptive-monitoring literature ([[syntheses/redesign-heuristics-vs-prpm|side-by-side mapping]]) covered ≈10/29 at runtime. APM extends the runtime surface further.

The result is **not a fresh catalogue replacing 2005** but a re-reading. Reijers & Liman Mansar's seven framework elements still hold; the [[concepts/devils-quadrangle|devil's quadrangle]] still scaffolds the trade-off conversation. What changes is (a) the **direction and magnitude** of each dimension, (b) the **failure modes** that previously did not exist, and (c) the **set of additional heuristics** the catalogue must absorb to remain complete.

## Three editorial moves

For each 2005 heuristic this synthesis assigns one of four reinterpretation tags:

| Tag | Meaning |
|---|---|
| **R — re-anchored** | The lever still pulls in the same direction. What changes is the *actor* (an agent, not a human) or the *substrate* (a frame, not a process diagram). Trade-off profile largely preserved. |
| **A — amplified** | The lever becomes structurally cheaper or faster to operate in the agent era. The 2005 trade-off bites less; a heuristic that was discretionary in 1995 is close to default in 2026. |
| **I — inverted / re-priced** | The trade-off direction *flips* on at least one axis, or new costs (alignment, hallucination, audit) appear that the 2005 four-axis quadrangle does not name. |
| **N — newly enacted at runtime** | The 2005 catalogue placed this heuristic at design time. In an agentic system the same lever is now pulled per-case by an agent under its frame. |

A heuristic may carry more than one tag (e.g., R + N: same direction, now run-time).

## The devil's quadrangle, re-defined for agents

The four axes survive but their **operational referents** shift:

| 2005 axis | Classical referent | Agent-era referent (2026) |
|---|---|---|
| **Time** | Cycle, lead, waiting time (human-bound) | Wall-clock + **agent latency** (model inference, tool round-trips, plan-revise loops); waiting time often collapses to milliseconds, but plan-revise can re-introduce it. |
| **Cost** | Operational cost per case (labour + tooling) | Labour + **token / inference cost** + tool-call cost + verification cost (a "judge" pass is now a meaningful line item). |
| **Quality** | External (customer-perceived) and internal (defect-free) | Adds **hallucination rate**, **frame-violation rate**, **alignment quality**, **explainability**; defects are now sometimes *plausible-but-wrong*, a class the 2005 quadrangle does not name. |
| **Flexibility** | Ability to handle variation / change | Variation handled by [[concepts/framed-autonomy|framed autonomy]]; new flexibility *primitives* (prompt rewrite, tool-set expansion, sub-agent spawning); but **frame churn** can degrade reproducibility — a flexibility cost. |

Two cross-cutting tensions appear that the 2005 four-axis model does not absorb cleanly and that this synthesis flags explicitly in every reinterpretation:

- **Auditability / verifiability.** Behaviour is now non-deterministic. A trade-off axis the quadrangle hides under "quality." See [[sources/2025-fournier-agentic-ai-process-observability|Fournier et al. (2025)]] on agent observability.
- **Safety / alignment.** Agents may take unintended actions inside their permission set. [[sources/2025-calvanese-autonomy-business-process-execution|Calvanese et al. (2025)]] treat the normative frame as the bearer of safety.

These are not new axes of the quadrangle in this synthesis — they are **dimension-of-quality refinements** the agent era forces into view. Future work may justify a five- or six-pointed *devil's polygon*; here we stay faithful to the 1995 scaffold and treat them as quality sub-axes.

---

## Part A — The 29 heuristics, reinterpreted

Each entry is structured: *element · tag(s) · agent-era reading · revised devil's-quadrangle profile (↑ improves the dimension, ↓ degrades it, — neutral) · new failure modes*.

### Customers (3)

**1. Control relocation** — *Customers · R, N*. Move controls to the customer side. In 2005 this meant self-service kiosks and customer-completed forms. In 2026 the customer's *own* agent (their personal assistant) negotiates with the firm's agent via [[concepts/mcp|MCP]]/ACP. Quality ↑ (customer's agent enforces the customer's preferences); cost ↓ (the firm pays neither for staff nor for kiosk hardware); time ↑ (round-trips are millisecond-cheap); flexibility ↑ (the customer's agent absorbs format variance). **New failure modes:** the firm now depends on an opposing optimisation — the customer's agent maximises *the customer's* objective, not theirs; gaming and prompt-injection become first-order risks.

**2. Contact reduction** — *Customers · I*. The 2005 heuristic told you to *cut* the number of touchpoints because each touchpoint was expensive. In 2026 contact is essentially free, 24/7, and conversational. The trade-off direction inverts: **maximise meaningful contact** at zero marginal cost, but **minimise contact cost per useful exchange** by routing trivial contact to agents and reserving humans for irreducible cases. Time ↑, cost ↓ for routine contact, quality variable (depends on hallucination rate), flexibility ↑. **New failure modes:** contact inflation (the agent invites itself into the customer's flow), over-personalisation, GDPR exposure on conversational logs.

**3. Integration** — *Customers · A*. JIT integration with the customer's workflow used to be a one-off engineering effort. With [[concepts/mcp|MCP]]/ACP it is the default integration mode. Cost ↓↓, time ↓, flexibility ↑↑, quality depends on tool reliability. **New failure modes:** transitive trust through tool chains; an agent that is integrated with eight tools inherits the worst guarantees of any of them.

### Products (0)

Reijers & Liman Mansar left this category empty in 2005. The agent era populates it for the first time, primarily because *information products* (recommendations, summaries, drafts) are now generated at point of use:

- **0a. Product personalisation at request time** *(new)* — the product is generated per-customer, per-request, from a generic substrate. See [[concepts/conversational-actionability]]. *Devil's quadrangle:* time ↑ (no inventory), cost ↓ for variants, quality ↑ on fit, flexibility ↑↑. **New failure modes:** unverifiable claims inside the product, IP attribution, training-data leakage.

### Business Process Operation view (5)

**4. Order types** — *Operation · A, N*. Splitting flows by order type used to be a routing rule maintained by analysts. Agents now infer order type from the case context and route per case. Time ↑, quality ↑ on fit, cost ↓↓ on rule maintenance, flexibility ↑. **New failure modes:** silent re-routing (the agent picks a variant that was never validated end-to-end).

**5. Task elimination** — *Operation · A, N*. [[sources/2025-fournier-agentic-ai-process-observability|Process-observability]] makes it trivial to see which steps add no value across runtime trajectories. Self-modifying agents can propose elimination candidates ([[sources/2025-elyasaf-self-modifying-abps|Elyasaf et al. 2025]], *evolution* mode). Time ↑↑, cost ↓↓, quality conditional on whether the eliminated step carried a hidden check, flexibility neutral. **New failure modes:** Chesterton's Fence — agents eliminate steps whose purpose was implicit; *load-bearing* checks vanish into the metric.

**6. Order-based work** — *Operation · A*. Batch processing was a workaround for human / machine setup cost. Agents have zero setup cost between cases; *order-based* is the natural mode. Time ↓↓, cost ↓, quality ↑, flexibility ↑. **New failure modes:** loss of throughput economies in genuinely batchable downstream resources (e.g., manufacturing).

**7. Triage** — *Operation · A, N*. The classic high-value example for predictive process monitoring; now it is *continuous, learned, multi-axis triage* on every case ([[concepts/prescriptive-process-monitoring]]). Cost ↓, time ↓, quality ↑, flexibility ↑. **New failure modes:** triage bias inherits from training data; protected-attribute proxies create discrimination risk (EU AI Act exposure).

**8. Task composition** — *Operation · I*. Composing small tasks into one large task was a 2005 win for hand-off elimination. With agents the *reverse* often pays: **decompose** monoliths into small, individually verifiable subtasks each handled by a specialist sub-agent under a verification pattern ([[sources/2026-dumas-agentic-bpms-pyramid|Dumas et al. 2026]] §4 *Verification* pattern). Direction flips on cost (decomposition adds verification calls), but quality ↑↑ and trustability ↑. **New failure modes:** over-decomposition produces orchestration overhead and context-window thrash.

### Business Process Behaviour view (4)

**9. Resequencing** — *Behaviour · A, N*. Already a runtime primitive in PrPM ([[syntheses/redesign-heuristics-vs-prpm]]); agents extend it to *planning at every step*. Time ↑, cost ↓, flexibility ↑↑, quality depends on plan stability. **New failure modes:** plan thrashing — successive small re-plans cancel each other out, producing zero net progress.

**10. Parallelism** — *Behaviour · A*. Spawning parallel sub-agents is a one-line operation. Buzacott's caveat (parallelism beats sequential only when individual job time dominates throughput) is largely dissolved by elastic compute. Time ↓↓, cost ↑ on inference, flexibility ↑, quality ↑ via diversity of attempts. **New failure modes:** cost runaway on naively parallel agent swarms; conflict-resolution overhead between siblings.

**11. Knock-out** — *Behaviour · A, N*. Ordering knock-out tests cheapest-first becomes *predicted-probability-first*: agents can call a cheap classifier on each candidate test and reorder. Time ↑, cost ↓. **New failure modes:** a fast cheap classifier may *itself* hallucinate the knock-out decision; calibration drift over time. Pair with a confidence threshold (see *Hallucination knock-out*, Part B).

**12. Exception** — *Behaviour · I*. Isolating exceptional cases to a separate flow was a 2005 win because exceptions were *expensive* for human handlers. Agents are often *equally* productive on exceptions and on happy-path cases — the marginal cost of handling a rare exception falls. The heuristic's premise weakens: **let agents absorb exceptions in-line** until volume / risk justifies escalation. Time ↑, cost ↓, flexibility ↑, quality conditional on adequacy of the agent's tail-coverage. **New failure modes:** silent tail-coverage degradation — the agent says "I handled it" when it actually failed in a way no human handler would have.

### Organization — structure (7)

**13. Order assignment** — *Org. structure · R, N*. Case ownership end-to-end now means assigning a dedicated **agent instance** to a case, with its own conversation history and frame. Time ↑, quality ↑, flexibility ↑. **New failure modes:** instance leakage between cases (context bleed); orphaned instances when a case stalls.

**14. Flexible assignment** — *Org. structure · A, N*. Late-binding of resource to task is essentially free; agents are spawnable on demand. **The 2005 trade-off (defer ↓ queue time, ↑ quality, ↑ flexibility) becomes the default mode**, not an alternative to be chosen. **New failure modes:** non-determinism — the same case routed at two moments hits two different agent versions; reproducibility costs.

**15. Centralization** — *Org. structure · A*. Centralised resource pools are the *native* shape of agent compute. Flexibility ↑, time ↑, cost ↓. **New failure modes:** single-vendor or single-model lock-in; correlated failures (one model regression takes down all agent-mediated processes simultaneously).

**16. Split responsibilities** — *Org. structure · R*. Clean responsibility boundaries between agents map naturally to multi-agent orchestration patterns ([[sources/2026-dumas-agentic-bpms-pyramid|Dumas 2026]] §4: Sequential, Routing, Managerial). Time ↑, quality ↑, flexibility ↑. **New failure modes:** inter-agent miscommunication and goal drift — each agent optimises locally, the system globally drifts.

**17. Customer teams** — *Org. structure · A*. A per-customer or per-segment "team" becomes a per-customer **agent persona** with retained memory of that customer's history. Cost ↓↓ (cheap to instantiate many personas), time ↑, flexibility ↑, quality ↑ on relationship continuity. **New failure modes:** persona drift; memory leakage between customers (the GDPR / DORA failure mode).

**18. Numerical involvement** — *Org. structure · I, N*. In 2005 you minimised the number of *people* touching a case. In 2026 you minimise the number of **actor-boundary crossings** (human↔agent, agent↔agent), because each crossing is a context-transfer cost and a verification surface. The optimisation direction is preserved; the *unit* changes. **New failure modes:** invisible context-transfer loss between agents that "look like" they share state but don't.

**19. Case manager** — *Org. structure · R, N*. The 2005 "case manager" archetype — one accountable face visible to the customer, even if not performing the work — maps directly onto a **conversational-actionability** layer ([[concepts/conversational-actionability]]): an agent that holds the customer-facing thread while routing work to specialist sub-agents underneath. Quality ↑, customer-perceived continuity ↑, cost ↓ vs. dedicated human case managers. **New failure modes:** accountability dissolution — when the case manager *is* an agent, regulatory questions about who is responsible become live (EU AI Act, GDPR Art. 22, see [[syntheses/legal-risk-mapping-ai-process-automation]]).

### Organization — population (4)

**20. Extra resources** — *Org. population · A, N*. Adding capacity at a bottleneck is an autoscaling call. Time ↑, flexibility ↑, cost ↑ proportional to inference volume. **New failure modes:** bottleneck moves invisibly to the verifier / human-in-the-loop step; capacity added at the wrong layer.

**21. Specialist–generalist** — *Org. population · R, A, N*. Cheap to instantiate both. The 2005 mutually exclusive choice (specialists for time, generalists for flexibility) dissolves — agents can be *both*, sequenced or composed per case. Time ↑, flexibility ↑↑. **New failure modes:** quality drift when a generalist agent over-extends into specialist territory.

**22. Empower** — *Org. population · R, I*. "Empower workers, eliminate sign-offs" was a 2005 win for time and motivation. In 2026 the analogue is *exactly* [[concepts/framed-autonomy|framed autonomy]] — but the 2005 win is **constrained** rather than amplified: agents are *too* easy to empower, and the marginal cost of a sign-off (a verifier-agent pass) is so low that the right move is often to **add** rather than remove control. Direction partially inverts: time ↑, cost neutral, **quality / safety ↑↑ with retained sign-off**. **New failure modes:** unframed empowerment leads to plausible-but-wrong actions taken at machine speed.

**23. Control addition** — *Org. population · A, N*. Adding a checkpoint used to mean adding a person and a delay. Adding an **agent-as-judge** verification step costs cents and milliseconds. Quality ↑↑, time ↑ marginal, cost ↑ small. **New failure modes:** judge-LLM systematic bias correlated with generator-LLM bias (same training distribution).

### Information (1)

**24. Buffering** — *Information · A*. Caching upstream so downstream doesn't refetch was the 2005 framing. In 2026 the analogues include the **context window**, **RAG retrieval**, and **prompt caching** — and the cost gradient is huge (cache hit ≈ 10% of cache miss in token cost). Time ↑, cost ↓↓. **New failure modes:** stale cache + plausible-sounding wrong answers; context-window poisoning.

### Technology (2)

**25. Task automation** — *Technology · R, I*. The 2005 catalogue assumed automation = removing human labour from *physical or transactional* tasks. The agent era extends automation to **cognitive / judgement** tasks. Direction is preserved on time and cost; quality is conditional on task structure — agents excel at high-context, low-precision tasks and underperform at precise, low-context ones. **New failure modes:** *automation bias* in human supervisors ([[concepts/automation-bias]]); deskilling of human reviewers.

**26. Integral business-process technology** — *Technology · R*. The 2005 win (workflow systems, shared databases, expert systems) is now an **A-BPMS platform** ([[sources/2026-dumas-agentic-bpms-pyramid|Dumas et al. 2026]]) — five layers, MCP integration, agentic orchestration. Time ↑, quality ↑, cost ↓ on lock-in horizon. **New failure modes:** platform monoculture; agent-platform vendor capture.

### External Environment (3)

**27. Trusted party** — *External · R*. Defer to a third party's judgement. Today the third party is often *another agent* — a credit-scoring API, a verifier service, a registry. Cost ↓, time ↑. **New failure modes:** opaque liability when the trusted party is an LLM API.

**28. Outsourcing** — *External · R*. Outsource non-core work — to an external agent fleet or to an AI-as-a-service provider. Cost ↓, quality conditional. **New failure modes:** training-data confidentiality leaks; jurisdiction-of-data exposure.

**29. Interfacing** — *External · A*. Standardising inter-organisational interfaces is now a protocol problem solved by [[concepts/mcp|MCP]]/ACP and shared ontologies. Cost ↓↓, time ↑, quality ↑. **New failure modes:** prompt-injection across organisational boundaries; protocol-level supply-chain attacks.

---

## Part B — New heuristics the 2005 catalogue does not contain

The agent era requires a small set of additions. Each is named, placed against the seven framework elements, and tagged against the devil's quadrangle. None of these has a 2005 predecessor.

### B1 — Frame-first design *(Organization — population / Organization — structure)*
Before assigning tasks, **specify the normative frame** — obligations, permissions, prohibitions ([[concepts/framed-autonomy]], [[sources/2025-calvanese-autonomy-business-process-execution]]). Without a frame, agent empowerment is reckless. *Devil's quadrangle:* quality ↑↑ (frame-violation ↓), flexibility ↑ (more autonomy is safe when framed), cost ↑ at design time, time ↑ at design time but ↓ at runtime. *Failure mode:* frame under-specification (the agent finds a path the frame did not forbid).

### B2 — Verification-by-judge *(Behaviour / Org. population)*
Pair every agent that produces consequential output with a **verifier agent** that scores it against the frame. Generalises *control addition* (#23) and makes it a default. *Devil's quadrangle:* quality ↑↑, time ↑ marginal, cost ↑ small, flexibility neutral. *Failure mode:* correlated bias between generator and judge; the judge under-detects exactly the errors the generator over-produces.

### B3 — Tool-mediated action *(Technology / Org. structure)*
Agents act **only through declared tools** with explicit scopes — never directly on production systems. Implements *least-privilege* and turns the action surface into an auditable inventory ([[sources/2026-dumas-agentic-bpms-pyramid|Dumas et al. 2026]] §3 *action layer*). *Devil's quadrangle:* quality ↑, cost ↑ (tool-build effort up-front), time ↑ at first deployment then ↓, flexibility ↑↑ on safe substitution. *Failure mode:* tool-set bloat; under-scoped tools that combine to authorise more than each does alone.

### B4 — Trajectory observability *(Information)*
Log every agent's perception → reasoning trace → action sequence as first-class events ([[sources/2025-fournier-agentic-ai-process-observability]]). Pre-condition for both conformance checking against the frame and offline evaluation. *Devil's quadrangle:* quality ↑↑ (debuggable), cost ↑ (storage + redaction), time neutral, flexibility ↑ (you can refactor what you can see). *Failure mode:* sensitive data in traces (GDPR exposure).

### B5 — Adaptation/evolution separation *(Org. structure / Org. population)*
Distinguish **ephemeral instance-level adaptation** from **persistent model-level evolution** ([[sources/2025-elyasaf-self-modifying-abps|Elyasaf et al. 2025]]). Adaptations that recur and prove safe are promoted into evolutions; evolutions that prove unsafe are rolled back. *Devil's quadrangle:* flexibility ↑↑, quality ↑ (governance gate), cost ↑ marginal, time ↑ on promotion review. *Failure mode:* adaptation buildup — many one-off patches accumulate without ever being promoted, creating drift between the canonical process and the lived one.

### B6 — Token / inference budget *(Information / External Environment)*
Cap **inference cost per case** and per session. Mirrors the operational-research idea of resource ceilings but applied to cognition. *Devil's quadrangle:* cost ↓↓ (predictable), time ↓ (caps run-away loops), quality ↓ on hard cases that need more thinking, flexibility neutral. *Failure mode:* the budget binds on exactly the cases that most need extra reasoning (high-value, edge cases).

### B7 — Confidence-thresholded knock-out *(Behaviour)*
Specialisation of *knock-out* (#11). When an agent's self-reported (or externally calibrated) confidence falls below θ, **stop and hand off** rather than commit. *Devil's quadrangle:* quality ↑↑, time ↑ at θ-crossings, cost ↑ on handoff. *Failure mode:* mis-calibrated confidence (LLMs are often confidently wrong); requires periodic re-calibration.

### B8 — Conversational handoff *(Customers / Org. structure)*
When the agent reaches its frame's boundary, **hand back conversationally** — to a human or to a more capable agent — without losing the customer-visible thread. Operationalises [[concepts/conversational-actionability]] under [[concepts/human-oversight|meaningful human oversight]]. *Devil's quadrangle:* quality ↑ (continuity), time ↑, cost ↑ (human time when invoked), flexibility ↑. *Failure mode:* premature handoff (every hard case to a human) or delayed handoff (the agent grinds past its competence).

### B9 — Frame-respecting outsourcing *(External Environment)*
Specialisation of *outsourcing* (#28) and *trusted party* (#27): when work crosses an organisational boundary to another agent, propagate the **frame** with it (deontic constraints, data-handling clauses), not only the task. *Devil's quadrangle:* cost ↑ at integration time, quality ↑↑, flexibility ↑. *Failure mode:* frame loss in transit; receiving party silently relaxes constraints.

### B10 — Determinism-by-request *(Behaviour / Information)*
For cases where reproducibility matters (regulatory, audit, recurrence), **pin** the agent's state — model version, seed, retrieval snapshot — so the same case yields the same trajectory. *Devil's quadrangle:* quality ↑ (auditability), time ↑ marginal, flexibility ↓ (you can't pick the best newer model mid-case), cost ↑ on snapshot retention. *Failure mode:* pinning to a deprecated model long after support has lapsed.

---

## Part C — Reading the catalogue against the devil's quadrangle (summary)

Counting tags across the 29 + 10 entries (39 total) by their **dominant** reinterpretation:

- **Amplified (A):** ≈14 — the largest class. The agent era makes cheap what was expensive: parallelism, triage, contact, buffering, late-binding, interfacing, customer teams, automation. The 2005 trade-offs become weaker; what was a careful decision becomes close to a default.
- **Re-anchored (R):** ≈11 — same direction, new actor. Case manager, split responsibilities, trusted party, outsourcing, automation, integral BP technology. These are continuities, not novelties.
- **Inverted / re-priced (I):** ≈5 — *contact reduction*, *task composition*, *exception isolation*, *empower*, *numerical involvement*. Each carries a trade-off direction that flips on at least one axis because the underlying cost structure changed.
- **Newly enacted at runtime (N):** ≈13 — heuristics now pulled per-case by an agent, often *in addition* to design-time enactment. Many of these were the same ones identified as PrPM analogues in the [[syntheses/redesign-heuristics-vs-prpm|2005-vs-PrPM mapping]] (#4, #5, #7, #9, #11, #14, #21, #23, etc.) — but APM widens the runtime surface beyond PrPM's case-level interventions.

The new heuristics (B1–B10) cluster heavily on **quality** improvements — frame, verification, observability, confidence, determinism — because the dominant agent-era trade-off is not time/cost (which the agents collapse) but **trust**. The devil's quadrangle still applies; **most agent-era trade-offs are paid in the *quality* axis**, especially in its hidden sub-axes (alignment, auditability, calibration, frame-respect).

## Part D — Three principles for using this catalogue

1. **The trade-off direction is now per-case, not per-process.** A heuristic that improves quality on the happy-path case may degrade it on a tail case (e.g., *exception in-line* #12). Always read the heuristic *in the context of the case distribution*, and where feasible let an [[concepts/prescriptive-process-monitoring|PrPM]] / agentic policy decide which heuristic to apply per case.
2. **Quality is the binding axis.** Where 2005's binding axis was usually time or cost, the agent era's binding axis is quality-as-trust. Plan the trade-off conversation around the quality sub-axes (hallucination, frame-violation, alignment, auditability) that 2005 left unnamed.
3. **The catalogue is a frame, not a script.** [[sources/2026-calvanese-agentic-bpm-manifesto|Calvanese et al. (2026)]]'s critique of activity-centric notation applies here too: the right unit of redesign is not "which heuristic do I apply" but **"what frame governs the agent, and which heuristics fall out of that frame as obligations vs. permissions?"** Heuristic *empower* (#22) becomes "what does the frame permit?"; *control addition* (#23) becomes "what does the frame oblige?"; *frame-first design* (B1) is the meta-heuristic that subsumes the rest.

---

## Practical artefacts

- **HTML overview map.** A self-contained pedagogical visualisation that places the (relevant) 2005 heuristics and the new B1–B10 against the four-axis devil's quadrangle, colour-coded by reinterpretation tag, with hover/expand reinterpretation notes: [[assets/ai-agent-heuristics-devils-quadrangle.html]]. The HTML's Part D carries a second-pass paper-angle brainstorm (six low-empiricism angles anchored in the Riess research arc + two horizon-expanding moves into formal-methods / philosophy-of-science adjacent communities); it replaces the first-pass brainstorm that was empirically heavy.
- **Side-by-side with PrPM.** The runtime-analogue gap analysis in [[syntheses/redesign-heuristics-vs-prpm]] is the natural companion: it identifies the ≈10/29 heuristics already enacted at runtime by PrPM; this synthesis widens that runtime surface to the full 29 + 10 under the APM frame.
- **BPM-evolution narrative.** Place this synthesis inside the seven-phase BPM arc of [[syntheses/bpm-phases-and-bpr-legacy|Phases and BPR legacy]] — it is the deliverable of Phase 6 (agentic era) against the Phase 1 anchor (BPR).

## Caveats and limits

- **Speculative where deployment data is thin.** [[sources/2025-vu-practitioner-perspectives-agent-governance|Vu et al. (2025)]] report that no studied practitioner had hands-on agentic BPM experience yet. Several of the "amplified" claims rest on architectural plausibility and prototype evidence rather than longitudinal field data. This synthesis should be revisited when post-2026 deployment case studies accumulate.
- **Technology-only scope, preserved from 2005.** The 2005 paper explicitly restricted scope to the *technical* challenge of redesign; this synthesis preserves that scope. The socio-cultural integration challenges of agentic BPM (job displacement, automation-bias-on-supervisors, accountability) are sketched only as *new failure modes* per heuristic. The broader socio-cultural agenda lives in [[syntheses/apm-business-themes]] and the legal-risk synthesis ([[syntheses/legal-risk-mapping-ai-process-automation]]).
- **Quadrangle pressure.** Several reinterpretations stretch the four-axis quadrangle (alignment, auditability, calibration). A future deliverable may justify a five- or six-pointed *devil's polygon*. For now we stay faithful to Brand & van der Kolk (1995) and Reijers & Liman Mansar (2005) and treat the new dimensions as quality sub-axes.

## Related

**Concepts:** [[concepts/bpr-heuristics]] · [[concepts/devils-quadrangle]] · [[concepts/agentic-bpm]] · [[concepts/framed-autonomy]] · [[concepts/self-modification]] · [[concepts/conversational-actionability]] · [[concepts/explainability-apm]] · [[concepts/perceive-reason-act]] · [[concepts/agentic-bpm-pyramid]] · [[concepts/prescriptive-process-monitoring]]

**Methods:** [[methods/process-redesign-heuristics]] · [[methods/agent-oriented-process-mining]]

**Sources:** [[sources/2005-reijers-limanmansar-best-practices-bpr]] · [[sources/2026-calvanese-agentic-bpm-manifesto]] · [[sources/2026-dumas-agentic-bpms-pyramid]] · [[sources/2025-calvanese-autonomy-business-process-execution]] · [[sources/2025-elyasaf-self-modifying-abps]] · [[sources/2025-fournier-agentic-ai-process-observability]] · [[sources/2025-vu-practitioner-perspectives-agent-governance]] · [[sources/2022-kubrak-prescriptive-ppm-slr]]

**Sibling syntheses:** [[syntheses/redesign-heuristics-vs-prpm]] (the PrPM-runtime mapping this work extends) · [[syntheses/bpm-phases-and-bpr-legacy]] (the seven-phase arc) · [[syntheses/apm-manifesto-core-messages]] · [[syntheses/apm-business-themes]] · [[syntheses/legal-risk-mapping-ai-process-automation]]