---
title: Self-Modification (Adaptation vs Evolution)
type: concept
tags: [apm, adaptation, evolution, self-adaptive, learning]
sources: ["[[sources/2026-calvanese-agentic-bpm-manifesto]]", "[[sources/2025-elyasaf-self-modifying-abps]]"]
created: 2026-04-13
updated: 2026-04-15
---

# Self-Modification

The fourth required capability of an APM agent ([[sources/2026-calvanese-agentic-bpm-manifesto]], §3.4): the ability of agents and agent coalitions to **adapt and evolve over time**. Self-modification is grounded in the agent's perception abilities and enables continuous improvement toward self-improving back-ends for organisations.

## Fundamental distinction: adaptation vs evolution

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

## Feedback cycle
Ideally, adaptation and evolution form a feedback cycle in the APM system: consistently effective adaptations become candidates for evolutionary incorporation into the process model — and thus into the agent's **frame** (see [[concepts/framed-autonomy]]).

Conversely, evolution should reduce the frequency of certain adaptations by pre-emptively addressing known failure modes.

## Design implications
- **Adaptation mechanisms** prioritise responsiveness, robustness, safety under uncertainty. Require runtime monitoring, exception handling, flexible execution engines.
- **Evolution mechanisms** require pattern mining across execution histories, causal inference, statistical validation, and change management protocols.

## Risks
- **Poisoning** — self-modification may evolve agents into unsafe behaviour if learning signals are corrupted.
- Uncertainty quantification and "learning-to-defer" approaches are needed to know when to return control to human agents.

## Elaboration in Elyasaf et al. 2025
[[sources/2025-elyasaf-self-modifying-abps]] operationalises the distinction with:
- A **5-dimensional classification** (adaptation/evolution; task/flow/process; reactive/proactive; human-driven/autonomous; planned/emergent).
- A **three-level autonomy roadmap** ([[concepts/abps-autonomy-levels]]).
- The **MAPE-K loop** ([[concepts/mape-k-loop]]) as the canonical adaptation architecture.
- Explicit research challenges on governance, continual learning, and [[concepts/aleatoric-vs-epistemic-uncertainty|uncertainty quantification]].

## Related
[[concepts/agentic-bpm]] · [[concepts/framed-autonomy]] · [[concepts/abps-autonomy-levels]] · [[concepts/mape-k-loop]] · [[concepts/agent-process-observability]] · [[methods/agent-oriented-process-mining]]