Designing Agentic AI Patterns: A Framework for Human-in-the-Loop Experiences
Designing agentic AI with no shared vocabulary
As agentic AI moved from research curiosity to product reality at AWS, a persistent problem surfaced: existing guidance was either too abstract or too narrow. High-level principles like "be transparent" and "keep users in control" articulate what should be achieved but leave open the question of how. Low-level UI patterns - confidence indicators, prompt fields, chat interfaces - prescribe implementation but don't help teams decide what to build in the first place.
What was missing was mid-level design knowledge: concepts specific enough to act on, general enough to apply across contexts. The absence was felt in practice. Traditional methods like user stories and user flows broke down when designing for non-deterministic AI. Teams struggled to frame problems, align on the right human-AI balance, and communicate decisions across disciplines.
Agentic AI also introduced a new coordination challenge. Unlike conventional software - where user actions produce predictable system reactions - agents can initiate actions, make decisions autonomously, and work out of sight. Getting the interplay between human and AI right is foundational to trust, safety, and usefulness. But no unified vocabulary existed for designing it.
Grounded in real products, validated in practice
I initiated this innovation workstream within AWS and secured funding to hire the first Design Scholar in the organisation - an HCI professor from the University of Washington - to work on this project. Working with the resulting cross-disciplinary team, which also included a generative AI product lead, senior engineers, and UX designers, I directed the design and research strategy from conception to publication.
We conducted landscape and artifact analysis of 60 commercially available AI applications, ranging from OpenAI's Deep Research and Adobe Photoshop to lesser-known tools with novel interaction models. We deliberately looked beyond dominant chat-interface patterns to surface emerging approaches to AI prominence, autonomy, and user involvement.
The framework was validated iteratively: through a formal design critique with six senior UX designers, reviews with scientists in responsible AI and open-source software, a presentation at Conflux - Amazon's flagship design conference with 6,000 attendees from all orgs within Amazon - and applied work across AWS SageMaker AI, where the framework is now in use across 200 projects. The result was a vocabulary that practitioners actually reached for in daily design work.
Three dimensions designers can directly control
We define human-AI coordination as the interplay of three dimensions. These map to the levers designers can actually manipulate - what is displayed, what users can do, and what the AI actually does. Other important qualities like trust, predictability, and agency are emergent: they arise from how these three are configured.
How prominently the AI presents itself and is perceived by users. More visible, vocal, transparent, or anthropomorphic AI has higher salience.
The degree and form of effort a user invests in directing or monitoring the AI - constrained and enabled by what the interface makes possible.
What the AI actually does, regardless of what is exposed to users. Critical for trust: users form mental models of AI behaviour that may not reflect reality.
When both involvement and salience are high, coordination is high - a collaborative, deeply in-the-loop experience. When both are low, coordination is minimal - automated and backgrounded. The right balance is not fixed; it shifts with the task, the user, and the moment.
Beyond binary: a spectrum of human-AI collaboration
Most discussions of agentic AI reflect binary thinking - the human either is or is not "in the loop." A key insight of our framework is that coordination exists on a spectrum. We articulate four zones, each representing a different calibration of involvement and salience.
User and AI work closely together across multiple phases - initiation, monitoring, updating, completion. AI salience and human involvement are both high. The user is very much in the loop. Suited to complex, high-stakes, or unfamiliar tasks.
Tasks are handled by AI with minimal user input. The user initiates and reviews; most of the work happens out of view. AI salience is concentrated at the beginning and end. The user is barely in the loop.
AI works in the background without announcing itself. The user may not register the assistance at all. Smart sorting, predictive text, and personalised navigation fall here. The AI delivers outcomes quickly and clearly. The user is implicitly in the loop.
AI acts without user awareness or involvement. While largely outside the scope of UX design, it matters: users form mental models of hidden AI activity, and perceptions of covert behaviour affect trust even when nothing untoward is happening.
Mapping how involvement rises and falls across a workflow
Because both agents and users can work independently, coordination cannot be static. A typical workflow moves through multiple zones: high involvement during initiation as users define goals and constraints; lower involvement during execution; a spike again at review and next steps.
We visualise these shifts as coordination curves - a variation of user journey mapping that plots human involvement and AI salience across six phases: initiating, monitoring, updating, completing, extending, and settings. High-level curves reveal the overall shape of an experience. Looking beneath the surface exposes specific AI touchpoints, handoffs, and decision points.
As agentic workflows grow longer and more computationally intensive, they create valleys in the coordination curve - stretches where the AI operates independently and the user is minimally involved. These valleys require thoughtful design around notification, approval, monitoring, and auditing. The UX layer must provide the transparency and controls needed to build trust and support course correction.
Four patterns that demonstrate the framework's generative capacity
Responsive salience
An AI agent that automatically adjusts its visibility and interaction intensity to match the context. When trust is low - a beginner user, a high-stakes task, sensitive data - the system increases salience: richer explanations, additional approval gates, expanded transparency. As confidence recovers, it quietly reverts. Early user testing validated the concept while revealing meaningful individual variation: some participants found high-salience modes exhausting, others valued the guidance. Several didn't notice the adaptation at all - a signal that when well calibrated, dynamic coordination can feel seamless rather than intrusive.
Workplan gating
Before executing a complex agentic task, the system surfaces a structured plan for user review and approval. This shifts the experience toward done-with-me at initiation, while allowing done-for-me execution once the user has signed off. It addresses a common user concern: wanting oversight of what the agent is planning before it acts, without having to approve every individual step.
Attribution markers
Micro-copy, placement, and iconography that make AI contribution legible without being intrusive. Markers calibrate salience at the content level - surfacing AI involvement where it matters for trust or accountability, keeping it quiet elsewhere. The pattern addresses the challenge of done-under-me experiences becoming covertly conducted without users realising it.
Progressive autonomy
A pattern for gradually extending AI autonomy as users build familiarity and trust with a system. Coordination starts in done-with-me territory and migrates toward done-for-me as the user's mental model matures. The shift can be user-initiated, system-suggested, or a combination - with clear affordances for dialling back if needed.
From internal knowledge base to published research
- Published on Amazon Science (March 2026), co-authored with a University of Washington HCI professor and an AWS AI product lead
- Academic paper submitted to ACM DIS 2026 - one of the field's leading venues for design and human-computer interaction research
- Framework adopted across all AWS SageMaker AI projects - now applied across 200 projects including production enterprise software used globally by thousands of users
- Prescriptive guidance site based on the framework distributed across AWS design and product organisations
- Presented at Conflux - Amazon's flagship design conference with 6,000 attendees from all orgs within Amazon - generating significant follow-up across design and engineering teams
- Validated through a formal design critique with multiple design and product teams and reviews with scientists in responsible AI
What this taught me about designing for non-determinism
The most important lesson from this work is that agentic AI demands a different design posture. With conventional software, the goal is to find the right static design. With agentic AI, the goal is to build systems and a shared vocabulary that evolve as we learn what works. The behaviour is unpredictable, so users and designers must adjust - but the technology itself can also learn, adapt, and course-correct proactively.
Building shared vocabulary matters as much as building the right interface. The framework's value wasn't just in the design patterns it generated - it was in giving cross-functional teams a common language for coordination decisions that had previously been implicit or contested. The coordination zones became shorthand that cut across design, engineering, product, and leadership conversations.
Teams that get this right won't simply build more capable agents. They will build agents that people trust, adopt, and find genuinely worth collaborating with.