Voices of the weekIdeas worth spending time with.

What keeps a team thinking together?

Sarah Wells

Independent technology consultant and LeadDev Editor at Large, working on platform engineering, AI governance and technical strategy · UK

The argument

Wells asks what happens to shared context, learning and challenge when engineers increasingly work with agents rather than with colleagues.

What is new

Her new editorial agenda puts the social work of engineering alongside questions about AI productivity and technical foundations.

Why it matters

Collaboration can develop judgement as well as produce an immediate result.

LecXie’s practical interpretationBefore removing a shared review, identify what people learn there and where that exchange will happen instead.

Evidence & limitsPractitioner observation and questions for investigation, not a measured decline in team learning. Her consultancy, publishing and speaking interests are visible. The research she discusses predates this week.

Make room for the learner’s judgement.

Jason Zagami

Senior Lecturer, School of Education and Professional Studies, Griffith University; researcher in educational technologies and teacher learning · Australia

The argument

Zagami examines a routine in which trainee teachers prompt an AI, evaluate its response and refine their work while documenting their reasoning.

What is new

Personalised agents are embedded in ordinary coursework, making the learning process—not merely the finished answer—the object of study.

Why it matters

Useful assistance should leave room to practise evaluation.

LecXie’s practical interpretationIn a low-stakes training exercise, ask learners to explain one suggestion they accepted and one they challenged. Do not turn private AI conversations into performance surveillance.

Evidence & limitsFourteen participants, one university, twelve weeks; exploratory qualitative evidence, not a causal learning-gain estimate. AI assisted analysis and writing. No relevant conflicts declared.

Support feels different when success looks zero-sum.

Bin Zhang, Yanlin Ma & Xiaolei Qian

Bin Zhang and Yanlin Ma: Business School, Sichuan University. Xiaolei Qian: Sichuan International Studies University and HELP University · China/Malaysia affiliations

The argument

Perceived AI support was associated with reflective strategies and reported initiative. The support–reflection relationship was weaker among respondents who saw organisational success as one person gaining at another’s expense.

What is new

The study examines a boundary condition: the beliefs people bring to an apparently supportive tool.

Why it matters

A friendly interface may coexist with reasons to withhold an improvement.

LecXie’s practical interpretationAsk whether sharing an AI-assisted improvement helps the team or threatens someone’s standing; test the answer locally rather than diagnosing individuals.

Evidence & limitsSelf-reports from 313 employees at six internet firms, collected in three waves. The zero-sum measure uses one item. Associations are not causal evidence; perceived support is not proof of AI goodwill. No funding or commercial conflicts declared.

LinkedIn profiles for these three authors have not been independently verified.

Check that the work happened.

Nigam Shah & Jonathan Chen

Shah is a Stanford Professor of Medicine and Chief Data Scientist at Stanford Health Care. Chen is a Stanford physician-scientist and Associate Professor in Computational Medicine and Biomedical Data Science · United States

The argument

Judge an agent by the completed workflow and the changes it makes in the working system, not just the quality of its answer.

What is new

Shah and Chen connect clinical and administrative benchmarks around failures between steps: information lost, actions omitted and work left unfinished.

Why it matters

An accurate description of an action is not evidence that it occurred.

LecXie’s practical interpretationDefine the observable end state before a trial. A service request should have the right owner, authorised action and recorded outcome—not merely a convincing completion message.

Evidence & limitsNew synthesis of earlier research. Selected tasks, simulated environments and specified models do not establish live hospital outcomes. The headline benchmark scores are not directly comparable. The authors are involved in the work discussed; Shah has company-founding interests, and the synthesis supplies no complete current interests declaration.

Can institutions cooperate under pressure?

Fundi Skweyiya-Gushu

Governance and public-policy researcher focused on AI governance, emerging technologies and international security; PhD candidate and former Armscor board member · South Africa

The argument

Responding to AI-enabled cybercrime requires institutions to practise sharing evidence and making decisions together, alongside formal agreements.

What is new

Skweyiya-Gushu proposes recurring African cyber-resilience exercises that expose operational failures, with access rules and independent oversight for shared intelligence.

Why it matters

Capable organisations can still fail at the handover.

LecXie’s practical interpretationRehearse one incident with security, the operational owner and an external provider. Record where information or authority becomes unclear, without using live sensitive data.

Evidence & limitsSource-linked policy analysis, not an evaluated intervention. Earlier operations and exercises inform the proposal; they are not new October results. Her former Armscor board role is disclosed.

Her LinkedIn address is sourced from a 2022 published CV; current profile access is unconfirmed.