Beyond Agentic Translation: A Design-Based Framework for Human-AI Collaboration in Translator Education

Synthesis: Wang (2026) proposes a design-based research framework reorienting translator education away from routine translation skills — increasingly performed end-to-end by agentic translation systems — toward four competencies: cross-disciplinary knowledge (D1), AI orchestration competence (D2), higher-order strategic judgment (D3), and critical AI evaluation (D4). The framework includes a translation-task displacement inventory mapping 18 sub-tasks to current AI capability and remaining human role, draft assessment rubrics, and a pre-registered 12-month DBR protocol (5-person co-design team, five pilot participants). As a protocol paper it reports no empirical data: the framework's effectiveness is untested and contingent on the planned pilot's results, and its transferability beyond a single Chinese institutional context is a hypothesis, not a finding.

Summary

Agentic translation systems — autonomous multi-agent architectures that plan, orchestrate tools, and produce end-to-end translations — increasingly perform the routine translation tasks translator-education programs were historically designed to teach. Without reform, the author argues, programs risk graduating students into occupational categories projected to be increasingly automated. This protocol paper responds with a framework grounded in distributed cognition, task-based automation theory, human-AI collaboration theory, AI literacy, and the European Master's in Translation (EMT) competence framework.

The framework's components are threefold. First, a translation-task displacement inventory maps 18 translation sub-tasks to current AI capability, displacement likelihood, and the remaining human role — each displacement-likelihood rating annotated [E] (supported by published empirical evidence) or [P] (based on the author's informed projection), a transparency convention that keeps speculation distinguishable from evidence. Second, a four-dimensional competence framework (D1–D4) with draft assessment rubrics redirects training from producing translations toward knowing when to delegate, how to orchestrate agents, when to intervene strategically, and how to critically evaluate output. Third, a 12-month design-based research protocol with pre-registered evaluation criteria on the Open Science Framework specifies a 5-person faculty co-design team and five separately recruited pilot participants, with implementation to be reported in a subsequent paper after ethical approval.

Key Findings

• Routine-skill displacement is the framework's premise, and the inventory makes it auditable. The 18-sub-task inventory rates each task's displacement likelihood with an [E]/[P] provenance tag — a reporting convention that separates evidence-backed claims from the author's projections and makes the framework's empirical foundation inspectable rather than asserted.
• Four competencies replace routine production as the training target. D1 cross-disciplinary knowledge (subject-matter expertise to direct and verify agentic output), D2 AI orchestration competence (decomposing tasks, sequencing agent actions, managing tool chains), D3 higher-order strategic judgment (deciding what to delegate and when to intervene), D4 critical AI evaluation (verifying quality, detecting errors, owning the result).
• The contribution is the method, not just the content. The framework's operationalization is explicitly developed through a co-design process rather than fixed by the paper; the reproducible small-team methodology and transferable protocol design are themselves offered as contributions, alongside the rubrics.
• Pre-registration and staged evidence. Evaluation criteria are pre-registered on OSF, and the paper's claims are staged: the framework is presented as a design foundation whose warrant depends on the pilot's results — no educational effectiveness is claimed in this paper.
• Grounded in existing competence frameworks rather than invented de novo. The design builds on the EMT framework's higher-order/strategic sub-competences, distributed cognition, and human-AI collaboration theory, positioning itself as an operationalization of established constructs for the agentic-AI era.

What this means for practice

• Program designers (translation and beyond). The displacement-inventory method transfers to any professional program whose routine tasks agentic systems increasingly perform: list the sub-tasks, rate displacement likelihood with provenance tags, and identify what human role remains — then redesign assessment around the remaining human work rather than the automated output.
• Instructors in AI-exposed disciplines. Treat AI orchestration as a teachable competence distinct from AI literacy: D2 covers decomposing tasks and sequencing agent actions, skills that "knowing how to use AI" framing typically omits.
• Assessment designers. The draft rubrics target delegation judgment (when to let the agent work vs. intervene) and critical evaluation of agentic output — competences invisible to assessments that score only the final translated product.
• Faculty developers. The 5-person co-design + pilot structure is a reproducible small-scale implementation pathway for curriculum reform where full-scale redesign is infeasible — the methodology is explicitly offered as transferable.
• Researchers evaluating the framework. The paper's honesty about what it does not establish (no empirical data, single-context, effectiveness untested) is the right model for reading any protocol paper: the framework is a hypothesis with a pre-registered test, not evidence.

Limitations

• No empirical data. This is a protocol paper: the framework's educational effectiveness is entirely untested, and the paper says so — its warrant is contingent on the planned pilot's results, and ethics approval for the baseline survey and pilot was still pending at writing.
• Single national and institutional context. All design work is set at one university in China; cross-cultural and cross-institutional transferability is explicitly a hypothesis requiring replication.
• Displacement ratings are part evidence, part projection. The [P]-tagged ratings rest on the author's informed judgment about where AI capability is heading, not on measurement of current systems; the inventory is a design instrument, not a capability audit.
• Co-design introduces the designers' priors. The framework's operationalization is developed through the co-design team, so the resulting rubrics and curriculum prototype embed the team's conceptions of competence — a known design-based-research limitation the author acknowledges alongside context-dependence.
• Longitudinal outcome questions unaddressed. Whether graduates trained under the framework achieve better career outcomes than conventionally trained peers is left to future follow-up.

Citation

Wang, J. Q. (2026). Beyond agentic translation: A design-based framework for human-AI collaboration in translator education. Education Sciences, 16(11), 1668. https://doi.org/10.3390/educsci16101668