Task progress/completion
+1.44Pre4/6
Post5.44/6
8 improved / 1 unchanged / 0 declined
MSc researchPilot evaluationTask-level agency
The Research
The MSc pilot evaluation behind Accai: a voice-first, multi-agent prototype for task initiation, focus, and return to work.
14/14
positive initiation responses
9
matched longer-session cases
8.33/10
average prototype rating
Research demo
A session as a support team, not a single chatbot
The prototype combines push-to-talk voice interaction, explicit agent handoffs, structured check-ins, and a persistent workspace so support can happen during work instead of only before it.
7-15 min
Voice-first entry and agent handoff
The system uses spoken interaction and explicit handoffs to translate intention into a first work block.
User goal
Article draft
Draft a 900-word article from a scattered note.
Grounding Guide
ActivationHold push-to-talk, name the article in one sentence, then open the note without fixing it yet.
Mid-block events can trigger brief prompts at halfway and final stretch.
Task-Agency Triangle
The research is organized around three task-agency signals
The report frames task-level agency as the immediate ability to start, continue, and re-engage with a chosen task under ordinary conditions of friction and distraction.
Task
Agency
Agency
Selected signal
Activation
Voice prompts, body-doubling-like presence, grounding, and a named first move.
14/14 live-session responses were positive for getting started, taking action, and knowing the next step.
Pilot evidence
Task-level agency moved in the intended direction
Experiment 2 compared matched baseline and end-session responses for nine participants after a 60-minute work session. The items were closely related rather than identical, so these are directional shifts rather than a strict repeated-measures scale.
Engagement/resumption
+1.00Pre4/6
Post5/6
5 improved / 4 unchanged / 0 declined
Phone/social distraction control
+1.44Pre3.89/6
Post5.33/6
6 improved / 3 unchanged / 0 declined
14/14
Live-session responses were positive for getting started, taking action, and knowing what to do next.
9/9
Participants rated voice clarity, not feeling rushed, and curiosity to explore more features positively.
6/9
Accessible and inclusive was the weakest usability item, pointing to interaction clarity as the main improvement area.
Mechanisms and friction
What users said made the support work
Mechanism
Activation support
Voice prompts helped participants get started when they felt stuck at the threshold into action.
Mechanism
Body-doubling-like presence
The agent was experienced as accompaniment, not only as instruction or task management.
Mechanism
Imaginal scaffolding
Early rehearsal prompts helped users visualize completing the task before beginning.
Friction
Workspace clarity
Left-side tabs and the purpose of the work area needed clearer signposting.
Friction
Active-agent legibility
Some participants wanted clearer cues about which voice or role was currently speaking.
Friction
Timing calibration
Spacebar control, prompt timing, and timer endpoints need refinement for smoother sessions.
Limits of inference
The report treats the findings as evidence of feasibility and mechanism rather than proven efficacy: the pilot was small, self-report heavy, and had no control condition.