Reinventing team flow within AI uncertainty

Avanscoperta Blog · article · 2026-05-14 source ↗ Reader 🖥️

Code Was Never the Bottleneck Measuring AI Impact Outcome-Based Work Design (vs. Fixed Jobs)Pair & Mob Programming Team Autonomy Token Throughput

Because large language models can touch nearly every step of software delivery at once rather than one isolated area, teams should respond to AI-driven uncertainty not with individual-level incentives or isolated experiments but with continuous, team-owned experimentation — resetting rituals, replacing estimation with throughput, and shifting from fixed roles to dynamic contributions — because the alternative (local, uncoordinated optimization) can make the whole system worse even when each individual change looks like a win.

practical · mixed — Diagnostic/expository in its systems-thinking analysis of how AI destabilizes team flow (Cynefin, Theory of Constraints, roles-vs-contributions), argumentative in its prescription: embrace continuous, team-owned experimentation (Popcorn Flow) instead of individual KPIs or one-off pilots.

Structure

Part	Role
Opening critique of management-imposed AI-usage KPIs and incentives (Tokenmaxxing)	context/counter-position — sets up what the author argues against
Small isolated experiments won't be enough / invalidating assumptions	diagnosis — AI's impact is systemic, not localized, so local experiments can combine to worsen the system
Emptied rituals; From Roles to Contributions; Bottlenecks are likely to move	diagnosis — names three specific team structures (ceremonies, roles, flow bottlenecks) destabilized by AI
Enter continuous experimentation (reset rituals, throughput instead of estimation, Popcorn Flow, assess learning as a team, repeat experiments, think out of the prompt, observe weak signals)	prescription — the concrete practice recommended to replace both top-down KPIs and uncoordinated individual experimentation

Key terms

Tokenmaxxing Token Throughput
Borrowed from Gergely Orosz: the pattern of organizations incentivizing AI usage by measuring token consumption, which the author argues is close to certain to be gamed and is especially unreliable in collaborative settings like pairing.
Cynefin (the complex/chaotic liminal space)
Used to characterize where AI tool adoption places a team's work: not in the 'complex' domain (where probing and sense-making still apply within some structure) but in a liminal space edging toward 'chaotic,' where there is no rulebook and the priority is bringing the team into 'more manageable waters' rather than following a procedure.
Roles vs. contributions
A proposed shift away from organizing work around fixed job titles/roles (which map to salary and specific responsibilities, and create handoffs and defensiveness when challenged) toward organizing around the dynamic contributions individuals actually make in a period where new AI-era skills don't track cleanly onto old seniority or title structures.
Popcorn Flow
Claudio Perrone's kanban-like board for continuous experimentation (states: Problem & Observations, Options, Possible Experiments, Committed, Ongoing, Review, Next), recommended here specifically because it lowers the threshold for starting an experiment to a 'why not?' level and builds in an explicit review step.
Throughput (as replacement for estimation) Measuring AI Impact
The number of items completed per unit time, proposed as the metric of choice once estimates become too variable to be useful under AI-driven uncertainty — valued because it preserves system observability and guards against local optimizations that clog the real bottleneck, without requiring speculative effort sizing.

Propositions

explicitManagement-imposed AI-usage KPIs and incentives for individuals and teams are likely to produce gamed or inconsistent results rather than genuine learning, because organizations tend to pick the wrong metric, and individual-level measures like token consumption break down in collaborative settings such as pairing and can also artificially narrow exploration if scoped to one vendor.conf: high
extendsAgentics / Tech Things: Tokenmaxxing is dead, long live tokenmaxxing · p2 — Adds a pairing-specific failure mode (individual token consumption is inconsistent to measure when two people share a session) to the general tokenmaxxing-as-gamed-incentive critique.
explicitUnlike most prior technical improvements, LLM tools have no single area of impact — they can simultaneously reshape requirements gathering, documentation, coding, design sparring, vulnerability checking, and non-critical specialist work — which means small, isolated experiments confined to one step risk missing (or actively worsening) systemic effects across the whole development flow.conf: high
explicitTwo independently reasonable local AI experiments can combine to make the overall system worse — e.g. a clean LLM-authored specification taken as-is and turned directly into a prompt by a developer, making a supposedly value-adding handoff redundant, or a faster bug-fix cycle that simply clogs an unchanged downstream pull-request approval queue — a Theory-of-Constraints-style local optimization that worsens whole-system performance.conf: high
supportsSoftware Factories, Light and Dark · p6 — Both describe the same mechanism from opposite ends: Osmani notes that removing the human review step makes perceived throughput jump; this piece gives the concrete downstream consequence — a faster fix cycle just clogs the unchanged pull-request approval queue, a local win that doesn't improve the system.
explicitIntroducing AI tools moves a team's day-to-day work from the 'complex' into a liminal space bordering 'chaotic' in Cynefin terms, especially once the impact on team collaboration (not just individual coding) is included — there is no detailed rulebook for this territory, so the task is to bring the team into more manageable waters, not to follow a known procedure.conf: high
explicitEstablished ceremonies and rituals (iteration length, planning meetings, retrospectives) were tuned for a pre-AI world — a different team size and interaction pattern — and are increasingly dissonant; they need to be explicitly reset and rebuilt to regain legitimacy rather than kept running on autopilot.conf: high
explicitFixed roles and job titles create friction (more handoffs, defensive protection of role boundaries, especially when tied to bonuses/KPIs) precisely when AI reshuffles which skills matter most; the more useful organizing unit in a period of change is the dynamic contribution an individual makes, because emerging AI-era skills (e.g. decomposing a complex feature into manageable steps) correlate more with attitude than with official seniority or title.conf: high
supportsFrom Async Code Reviews to Co-Creation Patterns · p11 — Both treat pairing/ensemble work as a team-level practice under pressure from AI-era changes to flow, rather than an individual productivity technique.
explicitA good measurement target for AI's systemic impact has skin in the game and real feedback loops — a team owning its own features in production is a much better target than a single contributor's delivery of a temporary artifact.conf: high
explicitcorroboratedAI making coding faster does not establish that coding is the system's bottleneck; the constraint is at least as likely to move downstream (approval queues backing up behind a faster fix cycle) or upstream (product roles sucked into continuous clarification loops with LLMs, or business validation failing to keep pace) as to disappear.conf: high
supports ←Software Teaming: A Mob Programming, Whole-Team Approach · p8 — The book's queueing/Lean Flow argument (coding was never the bottleneck; separation and handoffs were) is the pre-AI grounding for the claim that AI speeding up coding doesn't establish coding as the system's bottleneck.
explicitTraditional point estimation becomes close to worthless under AI-driven variability, so throughput is a better metric for tuning the pace of experimentation against strategic deadlines, because it protects system observability without speculative sizing effort and guards against local optimizations that clog the actual bottleneck.conf: high
explicitPopcorn Flow's board structure (Problem & Observations, Options, Possible Experiments, Committed, Ongoing, Review, Next) is well suited to this situation specifically because it lowers the threshold for starting an experiment to a 'why not?' level and builds an explicit review step into the format.conf: high
explicitThe consequences of an AI-driven experiment should be assessed by the team as a whole, not left to whoever introduced it, because an AI collaborator has no accountability and no skin in the game, so the team must supply the accountability an individual AI-paired-with-a-human experiment would otherwise lack.conf: high
explicitA single experiment run is not a valid test of whether an AI-era practice works, because tool versions, prompting techniques, and even small interaction details (e.g. dictation versus typing prompts) change fast enough that 'we tried this and it didn't work' is an unreliable conclusion — experiments must be repeated and tuned, and the resulting learning should itself be treated as ephemeral.conf: high
explicitBeyond pre-defined success criteria, teams should watch for weak and slow signals — including how people feel at the end of an AI-powered working day — because the cumulative and emotional effects of AI-driven change surface slowly and narrow, upfront success criteria won't capture unexpected outcomes.conf: high

Arguments

Teams should address AI's effects through continuous, team-owned experimentation rather than through individual-level KPIs/incentives or uncoordinated isolated pilots. strength: high

Explicit premises:

AI's impact is systemic, touching many SDLC steps simultaneously, rather than confined to one area.
Isolated local experiments — even individually reasonable ones — can combine via Theory-of-Constraints-style dynamics to make the whole system worse.
Individual-level incentives (Tokenmaxxing) are prone to gaming and break down in collaborative settings.

Implicit premises:

The team, not the individual who introduces a change, bears the downstream consequences of that change, so the team is the correct unit of both experimentation and accountability.

Inferences:

This implies that management's instinct to impose individual AI-usage KPIs and bonuses is not merely an imperfect proxy but actively counterproductive, since it measures and rewards exactly the wrong unit (the individual's token/output volume) for a systemic, team-level phenomenon.

Teams should reorganize around dynamic contributions rather than fixed roles during this period of AI-driven change. strength: medium