Gene Kim on Vibe Coding - Part II | Episode 411 | Troubleshooting Agile
Douglas Squirrel ยท video ยท 2025-12-18 source โ Reader ๐ฅ๏ธ
Calibrated Trust (in AI Agents)Changeability & OptionalityDORA AI Capabilities Model (Team Profiles & Capabilities)Jevons ParadoxSmall Teams & CompressionTrustVibe Coding
Vibe coding changes software teams' organizational unit of work โ toward smaller, option-rich teams and even individual business-stakeholder-plus-developer pairs rather than net headcount reduction โ and because trust in these tools is a learned skill that takes about a year to build, leaders should start adopting now rather than wait for the technology to mature further.
practical ยท mixed โ A conversational argument for action (restructure teams around vibe coding now, start adopting now) carried mostly through anecdote (the Toppo Pal story, the Carlos Baldwin call) rather than data or formal argument.
Structure
| Part | Role |
|---|---|
| Opening frame: 'should we fire all our engineers?' | sets up the organizational question this episode answers, in contrast to Part I's individual-productivity focus |
| Toppo Pal / Fidelity narrative (five-day solo build, team growth, junior engineer Swathy) | evidence against headcount reduction |
| Jevons's paradox applied to software production cost | theoretical frame generalizing the anecdote into a prediction |
| Team-size discussion (two-person team: problem-owner + fixer) | reframes the organizational unit of work |
| Option value / Carlos Baldwin call / nยทk/t/ฯ formula | theoretical frame for why vibe coding's economic effect runs through modularity, not just typing speed |
| Cost of delay / FedEx analogy for token spend | justifies treating high token spend as investment rather than waste |
| 'Is now the right time?' / trust-takes-12-months finding | practical conclusion: adopt now |
| Close: announcing Part III as a cross-examination | signals the hosts intend to stress-test these claims next episode |
Key terms
- option value
The economic value created by being able to run more parallel experiments (K) across more independent modules (N) with shorter cycle times (T) per experiment โ formalized loosely as a function of NยทK/T, with everything else (ฯ) absorbing the rest. Attributed to Carlos Baldwin's modularity research and the Amazon monolith-decomposition case. - cost of delay
The lost value from not having working software sooner; used to justify spending $500โ1000/day on AI tokens per engineer as equivalent to paying FedEx for overnight delivery. - trust (in AI tools) Trust
The ability to predict how the AI will act and react, which Kim claims rises roughly linearly over about 12 months of use and then plateaus around 15 months โ an unpublished finding from his DORA 2025 collaboration. - Jevons's paradox (applied to software production)
If the cost of producing software drops 10x, the result may be 10x more software (and more developer jobs) rather than 10x fewer developers โ extending the classical resource-efficiency paradox to software labor.
Propositions
- explicitcontestedVibe coding is unlikely to reduce engineering headcount, and in at least one observed case (the Toppo Pal / Fidelity application) it led to team growth and increased hiring, not reduction.conf: mediumcontradictsFragments: February 13 ยท p11 โ Kim argues vibe coding grows team size/headcount via latent demand; Fowler's fragments argues two-pizza teams stay roughly the same size and simply accomplish more โ both reject shrinkage, but for different reasons (growth vs. stasis).
- explicitIf the cost of software production drops by 10x, the outcome may be 10x more software produced rather than 10x fewer developers needed โ a Jevons's paradox effect applied to software labor.conf: mediumsupportsIs AI Doom Going Out of Style? ยท p3 โ Both apply Jevons's paradox to predict that cheaper software production expands total developer demand rather than shrinking it.
- explicitdevelopedVibe coding shifts the natural organizational unit toward a pair โ the person with the problem and the person who can fix it โ rather than a full cross-functional squad.conf: highsupportsYour Work Team Is Now a 'Pod' and Your Co-Workers Are AI Agents ยท p6 โ Both independently describe the smallest viable unit (problem-owner + fixer; Amazon's sub-two-pizza 'pods') shrinking below the traditional cross-functional squad as AI tooling improves.extends โGene Kim on Vibe Coding - Part III | Episode 412 | Troubleshooting Agile ยท p3 โ Part II's claim that vibe coding shifts the organizational unit to a pair (problem-owner + fixer) is extended here into the finer-grained atelier/task-graph model of how that pair (or a senior person plus many AI 'leaf-node' workers) actually divides labor.
- explicitVibe coding's deeper economic effect runs through increasing 'option value' โ more modules, more parallel experiments per module, and shorter cycle times per experiment โ the same mechanism Carlos Baldwin credits with the economic value unlocked by Amazon's service decomposition.conf: medium
- explicitBecause software's value is higher the sooner it arrives, organizations should be willing to pay a premium in AI token spend (framed as $500โ1000/engineer/day) to get the same code faster โ the cost-of-delay logic applied to compute spend.conf: medium
- explicitTrust in AI coding tools is a learned skill that rises roughly linearly over about 12 months of use and begins to plateau around 15 months โ an unpublished finding from Kim's 2025 DORA collaboration.conf: lowextendsThe Hidden Realities of AI Adoption ยท p9 โ The HBR roundtable frames trust as a leading indicator organizations should track over time; Kim's unpublished DORA finding supplies a concrete (if undocumented) timescale โ ~12 months โ for how that trust accrues.
- explicitBecause the trust curve takes about a year and the technology keeps improving regardless, teams should start using agentic coding tools now rather than deferring adoption until the tools mature further.conf: medium
- explicitAs trust in an AI tool rises, engineers give instructions with fewer words, need less supervision, and are willing to delegate bigger tasks to it.conf: medium
- implicitThe willingness or unwillingness to build and own AI-assisted software โ not raw coding skill โ is becoming the thing that determines who captures the value created, as shown by senior engineers declining to maintain an application a junior engineer then took over and was rewarded for.conf: medium
Arguments
Vibe coding will not produce a net reduction in engineering headcount, and may increase it. strength: low
Explicit premises:
- At Fidelity, a solo-built application created roughly 10x the expected internal demand and led to additional headcount rather than layoffs.
- Historically, agile and CI/CD reshaped team structure and operations without shrinking total engineering employment.
Implicit premises:
- Demand for software capability is latent and effectively unmet, not fixed, so cheaper production reveals new demand rather than just satisfying existing demand with fewer people.
- One organization's anecdote generalizes to the labor market as a whole.
Inferences:
- Hiring freezes or layoffs premised on 'AI will let us need fewer engineers' may be miscalibrated if this pattern generalizes.
Organizations and individuals should start adopting agentic/vibe coding tools now rather than waiting for them to mature. strength: low
Explicit premises:
- Trust in a nondeterministic AI tool is a learned skill requiring roughly 12 months of use to build, per Kim's unpublished DORA finding.
- The 'work deferral trap' (waiting for traffic to die down) means delay doesn't reduce the eventual time cost โ you still arrive later.
Implicit premises:
- The 12-month trust-building curve observed in Kim's (unspecified) sample generalizes across practitioners, tools, and contexts.
- There is no catastrophic downside to early adoption severe enough to outweigh the benefit of an earlier start on the trust curve.
Inferences:
- Organizations that delay adoption fall permanently behind on the trust/skill curve relative to early adopters, compounding over time.
Narratives
- Dr. Toppo Pal (VP of architecture at Fidelity, ex-Capital One DevOps leader) grew frustrated waiting on his team's six-month, two-hire estimate, coded a Neo4j-backed presentation layer himself in five days, found it drew ~10x the expected internal audience, watched senior engineers decline to maintain it, and handed it to the most junior engineer (Swathy), who then began out-earning her peers โ while overall headcount on the team grew rather than shrank.
โ Primary evidence for the headcount-growth and status-disruption propositions - Kim's Zoom call with Harvard Business School's Carlos Baldwin (modularity researcher, adviser to Steven Spear, inspiration for work connecting modularity to software architecture), discussing the nยทk/t/ฯ option-value formula and how Amazon's 1999โ2001 monolith decomposition created, by Kim's back-of-envelope estimate, a thousand- to ten-thousand-fold increase in option value โ which Kim credits as a contributing cause of AWS.
โ Supports the claim that vibe coding's real economic lever is option value via modularity, not generation speed alone - The FedEx overnight-shipping analogy: paying extra for a package to arrive today instead of next week is the same economic logic as paying $500โ1000/day in AI tokens to get code that would otherwise take two weeks delivered this afternoon.
โ Supports treating high token spend as a cost-of-delay investment rather than waste
Questions addressed
- โ
Should organizations fire engineers now that AI/vibe coding can write code?
No โ demand for software capability is effectively latent/unmet, so cheaper production tends to reveal more demand (and more headcount), as in the Toppo Pal case, rather than reducing the need for engineers. - โ
Is now the right time to adopt agentic/vibe coding, or is it better to wait for the tools to mature?
Start now โ trust in these tools takes roughly a year to build regardless of how capable the models become, so delaying adoption doesn't avoid that cost, it just pushes it later (the 'work deferral trap'). - โญ What specifically about vibe coding drives the claimed option-value gains โ the speed of code generation itself, or the modular decomposition it happens to encourage?
Gaps
- incompleteThe headcount-growth claim rests on a single self-reported anecdote (Toppo Pal / Fidelity) generalized into "team sizes aren't shrinking, they're growing" without any quantitative labor-market data, and without engaging the many documented counter-cases elsewhere in this corpus of shrinking teams, hiring freezes, and headcount cuts attributed to AI coding tools.
Read next: extract-dora-2025-state-of-ai-assisted-software-development (quantitative); extract-wsj-work-team-pod-ai-agents and extract-mosseri-ai-tailwind-authenticity (team-shrinkage claims) for the opposing data points. - uninformedThe '12 months to build trust in AI, plateauing at 15 months' finding is presented as coming from Kim's 2025 DORA collaboration but explicitly 'didn't make it into the report' โ no sample size, measurement method, or population is given, so the claim can't be checked.
Read next: Look for a published version of this finding, or a primary interview with Kim/DORA researchers describing the methodology. - incompleteThis is Part II of a conversation explicitly building on 'Part I' (individual-level productivity claims) and anticipates a Part III where Squirrel will cross-examine Kim's optimism โ neither is yet in this corpus, so the organizational claims here can't be checked against the individual-level foundation they rest on, or the stress-test that follows.
Read next: Extract Gene Kim on Vibe Coding Part I and Part III (Episode ~412/413, Troubleshooting Agile) when available.closed byGene Kim on Vibe Coding - Part III | Episode 412 | Troubleshooting Agile โ This is the promised Part III: Squirrel does cross-examine Kim's optimism (atelier/senior-talent bottleneck, the sawzall misuse risk, and acknowledged tool limitations), so the anticipated stress-test now exists in the corpus, though Part I (individual-level productivity claims) is still missing. - illogicalThe Jevons's-paradox-style prediction ("10x cheaper production โ 10x more software, more jobs") does not engage the fixed downstream-capacity constraint (review, testing, integration, compliance) that this corpus's own DORA 2025 extraction identifies as a hard limit on how much AI-generated throughput an organization can actually absorb โ nor the token-cost side of the same trade (the very $500โ1000/day figure Kim cites as evidence of enthusiasm could equally be read as a resource constraint on the demand-elasticity story).
Read next: extract-dora-2025-state-of-ai-assisted-software-development#p11
Related documents
- The Hidden Realities of AI Adoption
- Is AI Doom Going Out of Style?
- Your Work Team Is Now a 'Pod' and Your Co-Workers Are AI Agents
- Fragments: February 13
- AI Enthusiasts Are in a Race Against Time, AI Skeptics Are in a Race Against Entropy
- AI Is Changing the Rules of Entrepreneurship
- Gene Kim on Vibe Coding - Part III | Episode 412 | Troubleshooting Agile
- Your Company Needs an Energy Strategy for AI's Next Phase
- 11) AI-spesial med Vilde Opsal og Anders Sveen
- Leading your engineers towards an AI-assisted future