AI + Teams = ?

Marcus Castenfors · article · 2026-09-26 source ↗ Reader 🖥️

practical · mixed — A short Substack essay that reports three findings from an NYU 'buildathon' workshop study and converts each into concrete team practices to adopt — expository summary of someone else's research plus the author's own argued recommendations.

Structure

Key terms

• Collective debt
  The accruing misalignment cost of working individually with AI in parallel — progress feels fast and productive to the individual, but the group's shared sense of direction erodes and must eventually be repaid by reconciling divergent outputs.
• Single-player vs. multi-player AI usage
  Single-player: one person privately prompting AI alone. Multi-player: a team prompting together in the same place/session so the AI's output is immediately shared and debated rather than reconciled after the fact.
• Lean forward vs. lean back AI
  Lean forward: a person actively drives and AI supports (appropriate when defining the problem). Lean back: a person lets AI do more of the work (appropriate once solutioning, after the problem is clear).
• Idea vs. execution
  An 'idea' is the insight/product-discovery moment of correctly framing which problem to solve, drawn from synthesizing many (often tacit) data sources — a fundamentally human act. 'Execution' is building/prototyping/coding once the problem is clear — where AI excels.

Propositions

• explicitTeams that default to parallel, individual AI-chat work (everyone prompting alone) run into coordination trouble — information overload and wasted effort — because the speed of solo AI use skips the whiteboard-style alignment step teams used to take before splitting up work, and the resulting misalignment accrues as a 'debt to the collective.'conf: high
  supportsTBM 418: Campfires, Trails, and Quests · p5 — Both describe solo/parallel AI use as feeling productive while quietly eroding shared ground — this piece's 'collective debt' names the same phenomenon Cutler calls 'accelerated isolation that feels like collaboration.'
  supportsTrust Factory · p14 — Both locate the problem in single-player usage patterns specifically, not in the AI tool itself — this piece's workshop evidence corroborates Beck's claim that 'single-player usage' (not the genie) is what erodes team trust/alignment.
• implicitAI amplifies individual productivity while eroding a team's shared sense of direction, and the historical pattern of standing together to weigh options before splitting into individual work tends to get skipped specifically because AI makes solo progress feel fast and sufficient on its own.conf: high
• explicitAI is strong at execution (prototyping, coding, visualizing) but weak at problem-framing — generating the initial 'idea' or correctly identifying the right problem to solve — because that requires synthesizing across many, often tacit, human data sources, which is a fundamentally human act; the study's winning teams were distinguished by insight-led, collaborative problem definition, with AI providing only the execution engine.conf: high
• explicitThe best-performing teams in the study were not the most technically skilled; they were distinguished by diversity of perspective — someone who understood the policy landscape, someone with lived experience of the problem, someone who could reason about business models — which predicted innovation quality more than AI or technical proficiency did.conf: high
• explicitcontestedBecause AI now removes the technical barrier to prototyping, the standard product-team norm of the last two decades (a small, durable team of PM, designer, and a few engineers) should be revisited — teams should consider staffing or convening outside domain experts (lawyers, marketers, people with lived experience) directly into the work, since those experts can now prototype their own perspective rather than only being consulted and translated by engineers.conf: high
  contradicts ←Revised rules of engineering leadership. · p8 — Castenfors argues the standard durable small team's boundary should dissolve to let outside domain experts prototype directly; Larson argues durable, stably-staffed engineering teams become even more important as the fundamental unit — opposite prescriptions for team composition under the same AI-acceleration premise.
• explicitcorroboratedTeams should explicitly and deliberately define, as a written team norm, when AI use should be single-player versus multi-player, and when it should be lean-forward versus lean-back, rather than letting these defaults emerge implicitly from habit.conf: high
  supports ←John Cutler's Post · p5 — Castenfors concludes teams should explicitly, deliberately decide when AI use should be single-player vs. multiplayer rather than letting it default; Cutler's post supplies the sharper three-way vocabulary (pure single-player / shared-context solo / multiplayer) that such a deliberate team decision would need to use.

Arguments

Narratives

• The opening vignette personifies a new AI teammate as 'chatty, diligent... always cheerful and eager to help. Works around the clock. Doesn't eat lunch... after "thought for 42s", sometimes hallucinates and goes off on a tangent.'
  → Frames AI as a social/team-dynamics question rather than a tooling question from the first line, and foreshadows the piece's conclusion that AI's quirks (unreliable judgment, no shared context with the rest of the team) are precisely why deliberate team practices — not just better prompting — are what determines the outcome.
• The NYU 'buildathon': a 6-hour workshop where student teams tackled real New York City civic problems (grocery affordability, bike lane safety, childcare access) using AI tools at every stage, judged on solution quality and how effectively they used AI, with researchers observing, interviewing, and assessing the teams throughout.
  → Supplies the empirical ground for all three themes; functions evidentially rather than as mere illustration — every subsequent claim in the piece traces back to an observation or quote from this workshop.

Questions addressed

• ✅ What happens to team dynamics when AI teammates are introduced?
  Three themes: individual parallel AI use damages alignment; AI is strong at execution but weak at problem-framing; diversity of thought (not technical skill) predicts innovation quality.
• ⭕ Is there a general rule for when AI use should be single-player versus multi-player, or lean-forward versus lean-back?
  No fixed rule is given — only a prompt for each team to 'facilitate a session' and write down its own principles — so how a team identifies the relevant inflection points in the moment is left unanswered.
• ⭕ Does widening a team to include more, more-heterogeneous outside experts risk reproducing the same parallel-work misalignment problem the article diagnosed for individual AI use?

Gaps

• incompleteThe evidentiary base is a single 6-hour student 'buildathon' on invented civic problems with ad hoc teams formed for the exercise — not a real, ongoing product team with history, roles, incentives, and accountability. The piece draws durable team-structure recommendations ('revisit norms of team structures,' 'consider staffing more permanent experts') from a one-day exercise without addressing whether the findings generalize to long-lived teams with sunk-cost dynamics and hierarchy that a one-day hackathon team doesn't have.
  Read next: Look for longitudinal studies or case studies of real, ongoing product teams adopting shared/multiplayer AI practices and outside-expert staffing, to see whether the buildathon findings hold up over time and under real organizational constraints.
• incompleteThe piece recommends bringing more outside experts into the room to add diversity of thought, but never addresses the coordination/alignment cost of doing so — the same 'parallel work creates problems' dynamic it diagnosed for individual AI chats could plausibly recur with a wider, more heterogeneous group of contributors who have different working styles, tools, and incomplete shared context.
  Read next: Literature on cross-functional team size, coordination cost, and the trade-off between diversity-of-perspective gains and communication-overhead losses as group heterogeneity increases.
• incompleteNo method is given for how a team recognizes, in the moment, which decisions are 'important inflection points' requiring multi-player prompting versus the many smaller decisions safe to delegate individually — the piece asserts teams should 'set aside time' to define this but offers no concrete heuristic or worked example.
  Read next: Pair/mob programming literature and decision-point frameworks (e.g., when to pair vs. work solo) that already address a structurally similar problem for human-only collaboration.
  closed bySoftware Teaming: A Mob Programming, Whole-Team Approach — The 'Pause for Clarity' chapter gives exactly the concrete, named heuristic that gap asked for — a short enumerable list of in-the-moment triggers (solution not working, can't articulate the idea, code getting messy, anyone requests clarity) for recognizing when a decision point needs full-team consultation rather than solo handling, for human-only collaboration.

Closes gaps elsewhere

• Genie Lessons: Nobody Wants Agents · g2 — Gives a concrete, workshop-tested sketch of multiplayer AI collaboration that the cited gap said Beck never provided: prompting together at defined inflection points (especially problem-shaping), explicit single-player-vs-multi-player team norms, and a lean-forward/lean-back distinction for how actively a human drives versus delegates. It remains partial — the sketch comes from a one-day student exercise, not real-time collaborative software or a sustained team, so it narrows rather than fully closes the gap.

Related documents

• Trust Factory
• TBM 418: Campfires, Trails, and Quests
• AI Won't Save the Kingdoms We Built
  Coordination Cost Multiplayer vs Single-Player AI
• John Cutler's Post
  Coordination Cost Multiplayer vs Single-Player AI
• Addy Osmani’s Post
  AI Employee vs Tool Framing Taste & Judgment
• What happens when the coding becomes the least interesting part of the work
  Coordination Cost Taste & Judgment
• Experience Mapping Matters More the Faster You Move
  Coordination Cost Multiplayer vs Single-Player AI
• The founder of Postman says you have to kill your existing org chart
  Coordination Cost Taste & Judgment
• Adam Mosseri: AI is a tailwind for authenticity
  Coordination Cost Taste & Judgment
• Notes from FOSE Europe
  Coordination Cost Taste & Judgment