Agents Over Bubbles

Ben Thompson (Stratechery) · article · 2026-03 source ↗ Reader 🖥️

Agent Harness Agentic AI AI Bubble Debate Compute & Energy Infrastructure Agency Moats & Defensibility

The emergence of AI agents — which integrate models with software harnesses and multiply the output of a small number of agentic humans — means the massive AI infrastructure investment is justified demand rather than speculative bubble, and positions integrated model-plus-harness companies (Anthropic, OpenAI) as the durable winners in the value chain.

practical · argumentative — The article argues a specific thesis — that AI is not in a bubble — by marshaling evidence from industry trends, economic logic, and the evolution of LLM paradigms. It is practical in that it draws investment and business-strategy conclusions.

Structure

Part	Role
Three LLM paradigms (ChatGPT, o1, Opus 4.5/agents)	Establishes the trajectory of capability improvements that compound to make agents possible
The decreased need for agency	Shows that agents reduce the number of humans needed to generate massive compute demand and economic impact
Enterprise economic imperatives	Argues that enterprises will cut workforces and rebuild with agents because the economics are irresistible
Agents and the AI value chain	Argues that agent quality depends on model-harness integration, not commoditized models alone
Closing paradox	Self-aware coda acknowledging that declaring 'no bubble' is historically the moment bubbles form

Key terms

agent / agentic workload Agentic AI
A software system where a harness directs a model, uses deterministic tools to verify results, and operates without continuous human involvement — distinct from chatbots or reasoning models
harness Agent Harness
The software layer that controls the model, manages tool use, and orchestrates agentic workflows (e.g., Claude Code). The harness is what differentiates agents from raw models.
agency (human) Agency
The initiative and willingness of a human to direct AI toward productive ends; Thompson argues agents reduce the number of humans who need this quality
integration vs. modular Moats & Defensibility
Borrowed from Clayton Christensen / Thompson's prior work: profits flow to integrated (differentiated) parts of a value chain, not modular (commoditized) parts
paradigm (LLM paradigm)
A distinct mode of LLM usage: (1) chatbot, (2) reasoning model, (3) agent — each with qualitatively different compute profiles and user relationships

Propositions

explicitcontestedAI infrastructure investment is justified demand, not a speculative bubble.conf: high
contradicts ←Everyone I recommend on AI is optimistic. Except Cory Doctorow. · p3 — Doctorow's 'the AI build-out is a bubble' claim directly contradicts extract-agents-over-bubbles' thesis that 'AI infrastructure investment is justified demand, not a speculative bubble.'
explicitcorroboratedAgents represent a qualitative break from chatbots and reasoning models because they integrate models with harnesses and deterministic tools to verify results autonomously.conf: high
supports ←Meta Deploys Unified AI Agents to Automate Performance Optimization at Hyperscale · p2 — Meta's production platform instantiates Thompson's definition of agents as models integrated with harnesses, structured tools, and encoded knowledge.
explicitdevelopedThe number of humans who need to actively wield AI for massive compute demand to manifest is small and shrinking.conf: high
extends ←How AI Swarms Are Disrupting Democracy · p1 — Extends the few-humans-wielding-many-agents leverage claim into the adversarial domain: single operators supervising hundreds of thousands of disinformation agents.
explicitcorroboratedEnterprises will cut workforces and replace human coordination layers with agents because the economic imperatives are irresistible.conf: high
refines ←Your AI strategy has a trust problem, not a tooling problem · p6 — Qualifies which humans agents can displace: coordination layers yes, but agents lack the agency to replace high-agency employees.
supports ←Your AI strategy has a trust problem, not a tooling problem · p5 — Agrees the middle-management coordination layer shrinks as agents distribute context, supporting Thompson's replacement thesis for that layer.
explicitcorroboratedModel-harness integration is the point of differentiation in the AI value chain, not the model alone.conf: high
refines ←Building AI Agents For The Enterprise · p4 — Relocates the point of differentiation from model-harness integration to the encoded organisational knowledge fed into the harness.
supports ←GitHub - musistudio/claude-code-router: Use Claude Code as the foundation for coding infrastructure, allowing you to decide how to interact with the model while enjoying updates from Anthropic. · p11 — By treating the model as swappable behind Claude Code's harness, ccr embodies the claim that model-harness integration (the harness), not the model, is the locus of value.
explicitcontestedAnthropic and OpenAI are positioned to be significantly more profitable than previously expected because they own the integration point.conf: high
contradicts ←Andreessen Horowitz's Post · p17 — This post claims labs stay in the 'engine room' and leave packaging value to companies building on top; the other source argues the opposite — that owning the model-harness integration point (i.e. the packaging layer) is exactly what will make Anthropic and OpenAI more profitable than expected.
explicitcontestedMicrosoft's launch of Copilot Cowork (built on Claude/Anthropic, not model-agnostic) is evidence that model commoditization has failed.conf: high
contradicts ←GitHub - musistudio/claude-code-router: Use Claude Code as the foundation for coding infrastructure, allowing you to decide how to interact with the model while enjoying updates from Anthropic. · p11 — ccr runs Claude Code's harness across cheap/open models, treating the model as swappable — technical-feasibility evidence for model commoditization, cutting against the 'Copilot Cowork built on Claude proves commoditization has failed' claim.
contradicts ←Hugging Face's Clem Delangue: Stop Comparing Engines to Cars · p3 — Thompson reads Copilot Cowork's Claude dependency as proof model commoditization failed; Delangue argues the open-model gap is a harness-optimization artifact, so commoditization is still viable.
implicitApple's strategy of licensing Gemini and building its own harness (Siri) may be insufficient for agent-quality experiences.conf: medium
implicitConsumer AI demand may be satisfied by 'good enough' commodity models, but enterprise AI demand requires integrated agents.conf: medium
explicit⚡ disputed & corroboratedCompanies that do not adopt agents will be structurally disadvantaged against AI-native competitors with smaller cost structures and growing capabilities.conf: high
contradicts ←Breaking: Autonomous Agents are a Shitshow · p6 — Marcus argues autonomous-agent deployment is recklessly premature on security grounds, directly opposing Thompson's adopt-agents-or-be-structurally-disadvantaged imperative.
supports ←Building AI Agents For The Enterprise · p5 — Compounding self-educating AI systems give earliest movers a cumulative advantage, reinforcing the structural disadvantage of non-adopters.
explicitThe coming wave of layoffs is partially genuine AI-driven restructuring, not merely post-COVID corrections.conf: medium

Arguments

We are not in an AI bubble — the massive capex investment is warranted. strength: high

Explicit premises:

Agents dramatically increase compute demand per task (multiple model calls, CPU for harness, tool use).
A small number of agentic humans can generate enormous compute demand by controlling multiple agents.
Enterprise economic incentives (cost savings + top-line growth) create irresistible pull toward agent adoption.

Implicit premises:

Demand that is driven by measurable enterprise ROI is fundamentally different from speculative demand.
The three paradigms represent a monotonically improving trajectory that will continue.

Inferences:

The combination of fewer humans needed, higher compute per task, and strong economic incentives means demand will exceed supply for the foreseeable future.

Anthropic and OpenAI are not overvalued because they own the integration point in the AI value chain. strength: medium

Explicit premises:

Agent quality depends on model-harness integration, not model alone.
Profits flow to integrated (differentiated) parts of value chains, not modular (commoditized) parts.
Microsoft's Copilot Cowork abandoned model-agnosticism in favor of Anthropic's integrated stack.

Implicit premises:

The integration advantage is durable — competitors cannot replicate it easily.
The value-chain structure from hardware/software applies analogously to model/harness.

Inferences:

If models require integrated harnesses to deliver agent-quality experiences, then the model layer is not commoditized, and integrated providers capture surplus value.

Enterprises will replace human coordination layers with agents, leading to significant workforce reductions. strength: medium

Explicit premises:

A small number of people drive most value in organizations; the rest form a coordination apparatus.
That apparatus carries massive coordination costs with diminishing returns.
Agents tilt toward pure acceleration without the drag of human coordination.

Implicit premises:

Human coordination costs are the primary drag on organizational productivity.
Agents can adequately perform the roles currently filled by 'human cogs' in the organizational machine.

Inferences:

Companies will cut more rather than less, rebuilding scale with agents, or face competitive disadvantage from AI-native entrants.

Narratives

The three LLM paradigms: ChatGPT (2022), o1 (2024), Opus 4.5 + Claude Code (late 2025)
→ Historical scaffold that makes the agent paradigm feel like inevitable progression rather than hype
Opus 4.5 released quietly in November 2025, but became transformative only when Claude Code harness was updated in December
→ Key evidence that the harness — not the model — is the differentiator, supporting the integration argument
Microsoft's pivot from model-agnostic Core AI strategy to Copilot Cowork built on Anthropic's integrated stack
→ Serves as empirical falsification of the 'models are commodities' thesis
Asus CFO Nick Wu downplaying the MacBook Neo as 'just content consumption'
→ Illustrates that consumers primarily consume rather than produce, setting up why enterprise (not consumer) is the real AI market
Horace Dediu's argument that Apple made 'The Most Brilliant Move in Corporate History' by not building its own model
→ Serves as the strongest counter-argument to Thompson's thesis, which he then systematically dismantles via the agent integration argument

Questions addressed

✅ Is the current AI investment a speculative bubble?
No — agents create justified demand through higher compute per task, fewer humans needed, and irresistible enterprise economics.
✅ Are AI models being commoditized, and if so, where will profits flow?
Models alone may commoditize, but agents require model-harness integration, which is differentiated. Profits flow to Anthropic and OpenAI as integration points.
✅ Why is AI compute demand exceeding supply despite limited consumer adoption?
Agents multiply compute per user, and a small number of agentic humans can drive enormous demand without requiring mass adoption.
⭕ Will AI primarily eliminate jobs or create new productivity?
Thompson argues both — enterprises will cut coordination layers and use agents to make remaining high-value employees more productive — but acknowledges 'this is going to get ugly.'
⭕ Is Apple's strategy of licensing models and building its own harness viable?
Thompson is skeptical but leaves open the possibility that consumers don't care about agent quality, in which case Apple's approach suffices.

Gaps

incompleteThe article asserts enterprise ROI from agents but provides no quantitative evidence — no case studies of actual agent-driven productivity gains, cost savings, or revenue growth. The argument rests on anticipated rather than demonstrated returns.
Read next: Look for empirical studies or case reports on agent-driven enterprise productivity (e.g., from Anthropic's or OpenAI's enterprise customers).
closed byBuilding AI Agents For The Enterprise — Supplies the missing quantitative enterprise-agent evidence the gap asks for: Anthropic customer case studies with numbers (L'Oreal 44k users/99.9% accuracy, Lyft 87% faster resolution, Rakuten biweekly releases and 97% fewer critical errors).
incompleteThe durability of model-harness integration as a moat is asserted but not deeply examined. Could open-source harnesses (e.g., LangChain, custom orchestrators) eventually commoditize the harness layer too?
Read next: Explore whether open-source agent frameworks are closing the gap with proprietary harnesses like Claude Code.
closed byHugging Face's Clem Delangue: Stop Comparing Engines to Cars — Directly addresses whether the harness moat can be commoditized: open models only look weaker because harnesses were optimized for proprietary APIs, implying open harness+model stacks can close the gap (p1, p3).
incompleteThompson's claim that agents can replace 'human cogs' in organizational coordination assumes agents can handle ambiguous, political, and relational tasks — not just execution tasks like coding. No evidence is offered for this.
Read next: Research on AI in organizational coordination, decision-making, and management tasks.
incompleteThe article focuses almost entirely on coding agents as the exemplar of the agent paradigm. Whether agents generalize to non-coding domains with equal effectiveness is unexamined.
Read next: Look for agent applications outside software engineering — legal, financial, scientific, operational.
closed byBuilding AI Agents For The Enterprise — Documents agent effectiveness outside coding — L'Oreal conversational analytics and Lyft customer support with quantified results — addressing whether agents generalize beyond software engineering.
incompleteThe social and political consequences of the 'impending wave of layoffs' are acknowledged but not analyzed. Thompson says 'I'm not advocating for this outcome, rather analyzing why it is probably going to happen' — but the systemic effects (inequality, political backlash, regulatory response) are left entirely unexplored.
Read next: Read Daron Acemoglu's work on automation and labor markets; also policy analyses of AI-driven displacement.
closed byThe loudest warning about AI and jobs yet — Surveys exactly the systemic policy response that gap flagged as unanalyzed — the 'We Must Act Now' economist statement plus concrete safety-net proposals (unemployment-insurance overhaul, wage insurance, a sovereign wealth fund) — and features Daron Acemoglu himself, the gap's own suggested reading direction, as a signatory now taking the disruption risk seriously.

Agents Over Bubbles

Structure

Key terms

Propositions

Arguments

Narratives

Questions addressed

Gaps

Related documents