AgentArch addresses a critical gap in current research by positioning agentic AI systems as a new class of software systems that require dedicated architectural principles, patterns, and evaluation methodologies.
The rapid emergence of large language model (LLM)-based agents is transforming artificial intelligence from standalone models into complex, interactive software systems involving AI agents, multi-agent ecosystems, and human-AI collaboration. However, current research remains largely model-centric, with limited attention to the architectural foundations required to design, analyse, and govern such systems.
Recent advances in LLMs have rapidly shifted the focus of AI from standalone models to agentic systems, where LLMs are embedded within iterative reasoning, tool usage, and decision-making loops. These systems can function as autonomous or semi-autonomous agents that plan, invoke tools, and collaborate with humans or other agents in dynamic environments. This paradigm shift has led to the emergence of AI agents, multi-agent ecosystems, and human-AI collaborative systems across domains such as software engineering, healthcare, finance, and digital services.
This shift reflects a transition from model-centric AI to system-centric AI, in which agents function as executable entities with state, behaviour, and interaction capabilities. Agentic systems are no longer merely collections of models, but a new class of software systems that require principled design beyond model-level optimisation. Despite this rapid progress, existing agent frameworks are often constructed in an ad hoc manner, lacking principled abstractions for modularity, coordination, observability, and control.
AgentArch focuses on three interconnected dimensions: (i) the architecture of individual AI agents, including reasoning pipelines, memory, and tool integration; (ii) multi-agent systems, including coordination, communication, and emergent behaviours; and (iii) human-agent collaboration, where agents support or influence human decision-making processes. The workshop emphasises core architectural concerns, including modularity, composability, observability, reliability, and trust.
From an architectural perspective, agentic systems raise new challenges. AI agents can be viewed as modular and composable units, analogous to services in microservice architectures; however, unlike traditional services, they exhibit non-deterministic behaviour, maintain internal reasoning states, and interact with tools through complex feedback loops. Multi-agent systems introduce distributed and decentralised architectures requiring support for coordination protocols and conflict resolution. Human-agent collaboration introduces socio-technical considerations, raising concerns around transparency, traceability, controllability, and bias.
The scope of this workshop aligns well with the ICSA 2027 theme: the enduring role of software architecture in an evolving landscape. By bringing together researchers and practitioners across disciplines, the workshop aims to identify key challenges, develop architectural principles, and establish a research agenda for architecting next-generation AI-driven systems.
The workshop focuses on the architectural foundations of agentic AI systems, covering a broad yet coherent set of topics across three key dimensions. We particularly encourage work that explicitly considers agentic systems from an architectural perspective, including design principles, reusable patterns, system abstractions, and evaluation methodologies.
The primary goal of the AgentArch workshop is to establish a foundation for understanding and designing agentic AI systems from a software architecture perspective.
A highly interactive full-day workshop combining invited keynote talks, paper presentations, and collaborative working sessions. The format emphasises active participation and collaboration, going beyond traditional paper-centric workshops to foster deeper engagement and community building.
| Session | Duration | Description |
|---|---|---|
| Session 1 | 45 min |
Opening and Keynote
Keynote talk from a leading researcher or practitioner highlighting key challenges and emerging trends in architecting agentic AI systems.
|
| Session 2 | 90 min |
Paper Presentations I
Authors of accepted papers present their work, followed by short Q&A sessions to stimulate discussion and cross-fertilisation of ideas.
|
| Session 3 | 60 min |
Thematic Discussion Session
Organised around key themes — architectural patterns for AI agents; coordination and reliability in multi-agent systems; human-agent collaboration and trust. Each theme is introduced by a moderator, followed by open discussion.
|
| Session 4 | — |
Lunch Break
|
| Session 5 | 90 min |
Paper Presentations II
Authors of accepted papers present their work, followed by short Q&A sessions to stimulate discussion and cross-fertilisation of ideas.
|
| Session 6 | 30 min |
Breakout Working Groups
Small groups collaboratively explore specific research questions — identifying key architectural challenges, proposing design principles or solution directions, and outlining open research problems.
|
| Session 7 | 30 min |
Panel Discussion
Each group presents their findings, followed by a panel discussion involving organisers, invited speakers, and participants.
|
| Session 8 | 15 min |
Future Agenda
Synthesis session to summarise key insights and define a research agenda for architecting agentic AI systems.
|
We solicit contributions across two tracks. All submissions must follow the IEEE Computer Society proceedings format. Accepted papers will be included in the ICSA 2027 Companion Proceedings and published in the IEEE Xplore Digital Library.
All papers are to be submitted electronically via the EasyChair submission system by the submission deadline and must not have been published before or be submitted for review elsewhere while under consideration at AgentArch.
Novel frameworks, methodologies, or empirical studies presenting original research on the architectural foundations of agentic AI systems.
Position papers, preliminary results, tool demonstrations, or experience reports exploring emerging ideas and open challenges.
Best Paper Award. The workshop includes a Best Paper Award to recognise outstanding contributions and encourage high-quality submissions, particularly from early-career researchers.
All deadlines are Anywhere on Earth (AoE).
| Milestone | Date |
|---|---|
| Submissions Deadline | Dec 20, 2026 |
| Acceptance Notification | Jan 19, 2027 |
| Camera Ready | Jan 29, 2027 |
| Workshop Date | To be announced |
The workshop date will be announced shortly. Follow the ICSA 2027 website for conference updates.
