Agentic Infused Software Ecosystem (AISE)
- Agentic Infused Software Ecosystem (AISE) is a socio-technical framework that integrates AI agents, human oversight, and co-designed programming tools to enable transparent intent and operation.
- The architecture of AISE is organized around pillars such as AI agents, programming abstractions, and runtime environments, which ensure explicit intents, discoverability, and structured control.
- AISE emphasizes interoperability, verification, and governance by embedding prompts, workflows, and audit trails as first-class artifacts to support sustainable and accountable software engineering.
Agentic Infused Software Ecosystem (AISE) denotes a software stack deliberately co-designed so that AI agents, programming abstractions and tools, and the execution environment operate as first-class components of software development and operation (Marron, 24 Feb 2026). In the literature, AISE is not treated as a narrow layer of coding assistance. It is framed as a broader socio-technical ecosystem in which human and agent actors work across areas, activities, and artefacts, with varying levels of AI agency under human control, while prompts, workflows, controls, and organizational routines themselves become part of the engineered object (Hoda, 22 Oct 2025, Feldt et al., 16 Apr 2026).
1. Definition and conceptual scope
AISE emerges from the claim that agents alone are insufficient. The central argument is that contemporary software environments were built for human programmers and then retrofitted for AI assistance, leaving intent, constraints, discoverability, specification, and safety only weakly exposed in forms that humans and agents can reliably use (Marron, 24 Feb 2026). On this view, AISE is an ecosystem-level response: it makes agent use, tool use, validation, and runtime control explicit and analyzable rather than ad hoc.
This scope is broader than autonomous coding. A foundational framing paper on agentic software engineering argues for a “whole of process” vision spanning Agentic SE Ethical Alignment, Agentic Requirements Engineering, Agentic Design, Agentic Development, and Agentic Operations, and describes the field as involving human and agent actors, activities, artefacts, and varying levels of AI agency under human control (Hoda, 22 Oct 2025). AISE therefore includes lifecycle coverage, human oversight structures, and cross-cutting socio-technical concerns rather than only code-generation workflows.
A complementary conceptual model expands the engineering object outward from executable code to semi-executable artifacts: prompts, workflows, controls, operating logic, and societal and institutional fit. A semi-executable artifact is defined as a software-related artifact that helps specify, coordinate, or constrain behavior, but whose enactment depends on interpretation by humans, probabilistic models, or both, rather than on fully deterministic machine execution alone (Feldt et al., 16 Apr 2026). Within AISE, this means that system prompts, context files, workflow rules, evaluation harnesses, and escalation procedures are treated as engineering artifacts, not as informal scaffolding.
2. Foundational architecture and pillars
The most explicit architectural definition of AISE is organized around three pillars (Marron, 24 Feb 2026).
| Pillar | Role in AISE | Representative concerns |
|---|---|---|
| AI agents | Intent-understanding and action-suggesting components | tool use, decomposition, modularity, structured outputs |
| Programming languages, APIs, and tools | Communication substrate for humans and agents | explicitness, typing, validation, discoverability |
| Runtime environment and broader software ecosystem | Controlled execution and interaction with the external world | discovery, sandboxing, information control, resilience |
The paper that introduces AISE also grounds these pillars in five recurring design concepts: Explicit intents and behaviors, Discoverability, Mechanize everything, First-class cooperation, and Failure safety and resilience (Marron, 24 Feb 2026). These concepts are intended to align agent planning, human specification, tool invocation, and runtime enforcement under a shared set of abstractions.
One concrete instantiation uses Bosque as the programming-language substrate, with explicit api and agent constructs, type aliases, invariants, event-log-aware preconditions, and holes for deferred work; Sundew as the mechanized validation system translating programs to SMT-decidable forms; BAPI as typed API and data interchange with multiple serializations and sensitive annotations; and Mint as an agentic HATEOAS-style runtime that exposes /actions, /actions/{endpoint}, and /search routes for progressive discovery and execution (Marron, 24 Feb 2026). This design treats language, contracts, discovery, and runtime policy as one integrated substrate.
A different but complementary diagnostic model is the Semi-Executable Stack, which organizes the expanded engineering object into six rings: Executable artifacts, Instructional artifacts, Orchestrated execution, Control systems, Operating logic, and Societal and institutional fit (Feldt et al., 16 Apr 2026). Within AISE, this model functions less as a runtime stack than as a way to locate where a contribution or failure primarily sits, and which adjacent rings it depends on.
3. Interoperability as ecosystem substrate
A central systems argument in the AISE literature is that collaborative agentic AI will scale only if heterogeneous ecosystems interoperate. A position paper on interoperability identifies growing fragmentation among A2A, MCP, ACP, ANP, AITP, LMOS, agents.json, and Agora, and argues that fragmentation produces lock-in, switching costs, and adoption barriers (Sharma et al., 25 May 2025). It explicitly rejects both a single unified protocol and brittle post-hoc translation layers, and instead proposes minimal interoperability standards.
That proposal is the Web of Agents, a minimal architectural foundation with four building blocks.
| Building block | Function | Web technologies mapped in the paper |
|---|---|---|
| Agent-to-agent messaging | request/response coordination | HTTP requests |
| Interaction interoperability | interface and capability understanding | API documentation |
| State management | short-term and long-term continuity | sessions + DB integration |
| Agent discovery | publication and lookup of agents | URLs, DNS, well-known paths |
The messaging baseline is HTTP, with GET suggested for retrieving information and POST for more complex interactions. The interaction layer centers on an interaction document describing interface requirements, capabilities, inputs, outputs, and communication conditions; it may be structured JSON or natural language, and need not obey a rigid universal schema (Sharma et al., 25 May 2025). State management reuses web session patterns such as cookies and unique session identifiers for short-term continuity, and standard database integration for persistent memory. Discovery relies on URLs, DNS, well-known paths, and a search-engine-like model in which agents publish capability descriptions on the web and discovery services crawl and index them.
For AISE, the direct implication is that agents should be exposed as web-addressable services with stable, inspectable contracts and standard discovery surfaces, while richer internal frameworks remain compatible behind that boundary. The same paper treats transport compatibility, discoverability, interface descriptiveness, and state continuity as layered properties of interoperability rather than as one monolithic protocol feature (Sharma et al., 25 May 2025).
4. Process model, roles, and socio-technical organization
The process view of AISE is explicitly socio-technical. A foundational framing paper structures agentic software engineering around Activities, Actors, Artefacts (AAA) and argues that whole-process coverage must include requirements, design, development, operations, ethical alignment, process optimization, documentation, project management, and adoption (Hoda, 22 Oct 2025). The same paper proposes the CRAFT values—Comprehensive, Responsible, Adaptive, Foundational, and Translational—as a normative guide for ecosystem design. In AISE terms, this places trust, accountability, communication, coordination, and adoption alongside technical capability.
A curriculum paper makes the operational division of labor especially explicit: “humans frame, specify and judge, and agents execute” (Gorsky, 31 May 2026). It defines the anatomy of one workflow turn with seven parts: intent, specification, context, plan, execution, verification, and audit trail. The same paper introduces the evolutionary spiral as the operational form of the co-evolution of intent and build: inside a turn, work is exploratory and high velocity; between turns, work is pinned down through commit points, updated context, and drift detection (Gorsky, 31 May 2026).
This process model is consistent with the claim that prompts and plans are not sufficient by themselves. Within AISE, the durable artifacts of a turn include the specification handed to the agent, the context deliberately assembled for it, the verification evidence used to accept or reject the result, and the audit trail that records the whole process. The curriculum paper states that “a workflow that cannot be inspected, replayed or evaluated after the fact is craft, while engineering is accountable and reproducible” (Gorsky, 31 May 2026). That principle effectively elevates replayability and evidence preservation to ecosystem requirements.
The six-ring Semi-Executable Stack sharpens this further by separating where behavior is encoded directly in code from where it is encoded in prompts, workflows, controls, operating logic, and institutional constraints (Feldt et al., 16 Apr 2026). In AISE, that means failures can originate in the wrong ring: a system may have competent generation but weak control systems, or strong workflows but poor organizational operating logic.
5. Verification, observability, trust, and control
A recurring claim across the literature is that the difficult part of trustworthy agentic software is not merely generation quality, but intent inference and disciplined validation. A software-engineering position paper states that the core difficulty is “the deciphering and clarification of developer intent”, and that specification inference lies at the heart of software maintenance and program repair (Roychoudhury, 24 Aug 2025). It argues that future workflows will include AI-based verification and validation (V&V) of AI-generated code, because higher automation increases the volume of generated code and therefore increases the need for systematic checking.
Process observability has been developed beyond simple success/failure metrics. A process-centric analysis paper introduces Graphectory, a graph representation of trajectories with temporal and structural relations, and defines six process metrics: Node Count, Temporal Edge Count, Loop Count, Average Loop Length, Structural Edge Count, and Structural Breadth (Liu et al., 2 Dec 2025). On 4000 trajectories of SWE-agent and OpenHands with four backbone LLMs on SWE-bench Verified, the paper reports that richer prompts or stronger LLMs produce more complex Graphectory, that resolved issues tend to follow coherent Localization–Patching–Validation strategies, and that even successful runs often remain inefficient (Liu et al., 2 Dec 2025). AISE therefore requires process telemetry, not just outcome telemetry.
A separate architectural paper formalizes governance and observability as cross-cutting layers. Its reference architecture separates cognition from execution using typed tool interfaces and an explicit control layer with planner or policy logic, state machines, retry and backoff logic, and circuit breakers, while governance and observability include RBAC, audit logs, tracing, evaluation, policy enforcement, and cost and rate limits (Alenezi, 11 Feb 2026). The same paper presents a generic agent loop bounded by a maximum number of steps and by budgets on tokens, execution time, tool invocations, and monetary cost, making bounded autonomy an explicit control mechanism (Alenezi, 11 Feb 2026).
In domain-sensitive settings, governance can be pushed into machine-consumable domain artifacts. A methodological paper proposes GROUNDING.md as a community-governed, field-scoped epistemic grounding document that encodes Hard Constraints and Convention Parameters and overrides lower-level contexts when validity is at stake (Palmblad et al., 23 Apr 2026). Within AISE, this is a specialized but important pattern: field-level validity rules become a highest-priority control artifact above project plans and method guidance.
6. Implementations and ecosystem exemplars
The literature contains several concrete exemplars of AISE-like systems. In scientific computing, Bohrium + SciMaster is presented as an infrastructure-and-ecosystem stack for agentic science at scale (Zhang et al., 23 Dec 2025). Bohrium converts scientific data, software, compute, and laboratory systems into agent-ready capabilities across Reading, Computing, and Experiment; a scientific intelligence substrate organizes reusable models, knowledge, and community assets; and SciMaster orchestrates long-horizon workflows with state, validation, trace capture, and governed execution. The system is demonstrated with eleven representative master agents and is reported to generate execution-grounded signals from real workloads at multi-million scale, while achieving orders-of-magnitude reductions in end-to-end scientific cycle time in representative workflows (Zhang et al., 23 Dec 2025). This is one of the clearest realized domain-specific AISE stacks.
For open collaborative environments, Synergy defines the Open Agentic Web as a decentralized digital ecosystem in which agents from different users, organizations, and runtimes can discover one another, negotiate task boundaries, and delegate work across open technical and social surfaces (Nie et al., 30 Mar 2026). It introduces the notion of an Agentic Citizen, grounded in Agentic-Web-Native Collaboration, Agent Identity and Personhood, and Lifelong Evolution. Architecturally, Synergy uses sessions as execution capsules, Cortex for child-session orchestration, Holos for profile and contact systems, Agora for repository-backed workspaces, and an experience-centered learning mechanism that improves future behavior by recalling rewarded trajectories (Nie et al., 30 Mar 2026). This extends AISE from enterprise or laboratory settings into persistent, collaborative, repository-aware agent ecosystems.
A domain-specific security blueprint shows how AISE principles can be embedded directly into operational infrastructure. An adaptive cybersecurity architecture organizes work into three layers—Influx and contextual sensing, Agentic AI core, and Response and enforcement—and distributes autonomous agents across cloud, API, mobile, edge, and identity layers (Olayinka et al., 25 Sep 2025). In native cloud simulations, the reported results include precision 0.91, recall 0.87, F1 0.89, 220 ms response latency for autonomous mitigation, and low (<10%) resource overhead (Olayinka et al., 25 Sep 2025). The relevance to AISE lies in the pattern: sensing, reasoning, policy adaptation, and enforcement are embedded into the ecosystem itself rather than added as an external analytic layer.
7. Limitations and open research problems
Despite the breadth of proposals, many AISE papers are still conceptual, diagnostic, or architectural rather than complete protocol or deployment specifications. The interoperability literature explicitly notes unresolved gaps: no formal wire protocol, no normative universal schema, under-specified identity and trust, loose semantic interoperability, operationally underspecified discovery, and no standard for cross-agent state handoff or memory portability (Sharma et al., 25 May 2025). Similar incompleteness appears in broader architectural work, where governance, multi-agent coordination protocols, and formal verification remain active rather than settled topics.
A community research agenda identifies six thematic areas that remain open for agentic software engineering: Governance, Software Engineering for Agents, Agents for Software Architecture, Quality and Evaluation, Sustainability, and Code (Taibi et al., 12 May 2026). The same agenda emphasizes AgentOps, hybrid agent-based architectures, quality metrics for agents, robustness of agent ecosystems, sustainability debt, and governance debt as unresolved research and practice issues (Taibi et al., 12 May 2026). These concerns indicate that AISE is not merely an integration challenge; it is also a problem of lifecycle discipline, quality theory, organizational redesign, and long-term sustainability.
Taken together, the literature presents AISE as a deliberate reorganization of software around agents, explicit contracts, verification, observability, and governance. Its current state is best described as a convergence of architectural principles and domain exemplars rather than a single standardized stack. The recurrent direction is clear: AISE becomes more credible as agents are made interoperable, typed, discoverable, stateful, verifiable, auditable, and embedded in socio-technical operating logic rather than treated as isolated model endpoints (Marron, 24 Feb 2026, Hoda, 22 Oct 2025).