Balancing Teacher and Student Agency: Co-Orchestration Tool Design Supporting Real-Time Dynamic Pairing

Synthesis: This paper maps how 17 teachers and 13 students want control distributed across a teacher-facing intelligent tutoring co-orchestration tool that dynamically pairs students for collaborative problem-solving. Using participatory speed dating (74 co-generated design ideas narrowed to 34 storyboarded hybrids, ranked by all 30 participants with verbalized reasoning), the authors build an agency design space along a teacher-dominated-to-student-dominated axis for seven orchestration challenges spanning before/during/after pairing. The headline finding is a temporal gradient: both stakeholder groups prefer strong teacher control in preparation and initiation, then progressively value hybrid and student control as the activity unfolds — full student control is never preferred by either group. Teachers and students align on hybrid control for content selection, monitoring, transitions, and evaluation, but diverge on pairing timing (teachers want more control; students want readiness signaling) and partner selection (the exact inverse — students trust teacher choice, teachers want student voice). The practical recommendation is stage-sensitive agency: structured teacher guidance early, progressively increasing student autonomy later, with bounded student proposals validated by teacher-facing analytics rather than free-form mid-activity override requests.

The theoretical contribution is treating teacher and student control as independent dimensions rather than a zero-sum trade-off (after Eshel & Kohavi 2003): hybrid control means both groups can exercise high control simultaneously. The empirical contribution is the design space itself — 32 of the 34 ideas mapped onto a challenge × agency grid, with Kendall's W agreement coefficients per group — which turns "balance agency" into a concrete set of ranked design options per decision point. Notably, several "obviously desirable" autonomy mechanisms (students rejecting partners, switching assignments) rank last for teachers and near-last for students, because both groups anticipate gaming and social rather than learning-driven choices. Conversely, "save progress and resume" is the strongest cross-group consensus — abrupt, teacher-timed stops were universally disliked. The design implications generalize beyond this tool: make upcoming actions visible to students (previews, countdowns, advance notice) to build agency grounded in anticipation rather than control; accept structured, bounded student input (partner constraints, readiness signals) rather than real-time override; and let system-generated collaboration data — not subjective ratings — anchor future pairing decisions, since both groups distrust rating-based systems (friends rate high; rivals rate low) while both trust objective statistics.

Key findings

• A temporal gradient in control preference. Both groups preferred stronger teacher control in pre-pairing preparation (C1: "quiz for pairing" 1.7 students / "pairing rules" 2.5 teachers) and increasingly favored hybrid or student control in during- and post-collaboration stages (C4 content, C5 monitoring, C6 transitions, C7 evaluation all showed green hybrid bubbles).
• Full student control was never preferred by either group — the rightmost column of the design space is uniformly yellow. Student-initiated requests to change partners or assignments ranked last or near-last for both stakeholders, viewed as disruptive and gameable.
• Cross-group alignment: hybrid control for content selection (teachers: "Teacher proposes, students decide" 1.6; students: "Students propose, teacher decides" 1.83 — both hybrid), automated collaboration support (NPC help 2.0/1.7; keyword detection 1.5 teachers), and progress-preserving transitions ("Save progress and resume" 1.6/1.3).
• The two divergences are mirror images. Timing: teachers preferred "Teacher decides" (2.1) over student-decided timing, fearing chaos and schedule disruption; students preferred readiness self-signaling ("When students are ready" 1.2) to avoid rushing. Partners: students ranked "Teacher decides" first (1.9), citing teachers' knowledge of skills and friendship-distraction risks; teachers preferred "Teacher assigns, student chooses role" (2.2) — each group wanting the other to hold more control here.
• Objective data over subjective ratings for evaluation. Students' top choice was teacher pairing based on collaboration statistics (1.4); teachers preferred mutual rating (1.5); both ranked "students rate" last. Distrust of ratings was explicit: "students might rate their friends really high to be paired again, even if they won't be productive" (S7).
• Teachers were the more unanimous group in most challenges (Kendall's W teachers 0.424–0.748; students 0.108–0.702, with lowest agreement on C5 productivity monitoring), suggesting teachers share a coherent classroom-management prior that student preferences lack.
• NPC/automated support filled a real gap. Both groups valued NPC help as a backup when teachers are unavailable (shy students, teacher bandwidth), though both flagged overuse and effectiveness concerns.

What this means for practice

• For developers of teacher-facing orchestration tools: don't offer a single control dial — offer stage-dependent defaults. Structure (teacher-led) for preparation and initiation; hybrid proposals for content, monitoring, transitions, and evaluation. The paper's design space is directly reusable as a feature-priority map.
• Design for "agency by anticipation": previews, countdowns, and advance notice of upcoming actions give students preparedness without handing them control — the consensus "save progress and resume" pattern (both groups' top choice) is the concrete instance.
• Accept bounded proposals, not free override. Student partner-constraints or readiness signals submitted before pairing, filtered through teacher-facing analytics, satisfy the student-voice demand teachers acknowledged while protecting classroom manageability.
• Ground future pairing in system data, not ratings. Both groups trusted objective collaboration statistics; both distrusted peer ratings as gameable. If ratings are collected, pair them with system data and teacher adjudication, never stand alone.
• For researchers: the PSD method (co-generate 74 ideas → curate 34 storyboards → rank + reason) is a transferable protocol for multi-stakeholder control questions in educational AI, and the agency spectrum × challenge-stage grid is a reusable visualization.

Limitations

• Small, self-selected sample (17 teachers, 13 students; Phase 2 n=25) recruited via snowballing — not representative of districts, and participant motivation may skew toward tooling enthusiasts.
• All contexts are U.S.-based; the study's math-algebra tutoring ecosystem (MATHia + peer tutoring + BKT) is one specific configuration, and conclusions may not transfer to non-STEM or non-ITS domains.
• Storyboard-based evaluation (PSD) measures stated preference for designs, not enacted behavior in real classrooms; the paper explicitly notes regions of the design space (e.g., C1 with full student control) remain unmapped.
• Qualitative clustering is single-study with three coders; quantitative rankings are ordinal means without formal significance testing across groups — the alignment/divergence claims are descriptive.
• Audio-transcript accuracy of the Zoom caption pipeline was manually corrected, so coding depends on cleaned transcripts.

Citation

Yang, K. B., Liu, M., Xu, L., Rummel, N., & Aleven, V. (2026). Balancing Teacher and Student Agency: Co-Orchestration Tool Design Supporting Real-Time Dynamic Pairing. Proceedings of the ACM on Human-Computer Interaction, 10(6), Article CSCW054.