Agentic AI could remake firms: by shifting coordination from pairwise integration to standardized protocols, integration costs fall and an ‘‘hourglass’’ or 'Headless Firm' equilibrium emerges—thin protocol orchestrators over markets of micro-specialized agents—pushing unbundling in high knowledge-velocity sectors.

Main Finding

Agentic (tool-using, goal-directed) AI can change how coordination costs scale inside socio-technical systems. If integration is collapsed into a standardized protocol waist and verification work is reusable (i.e., per-task verification grows sublinearly with workflow width), total coordination cost shifts from topology-dominated (∝ number of pairwise edges, O(n^2)) to throughput-dominated (∝ tasks, O(T) and linear in n). That scaling change selects an “hourglass” organizational equilibrium—the Headless Firm—featuring: a personalized generative UI (intent layer) at the top, a thin protocol-and-policy waist, and a competitive market of micro-specialized execution agents below. This equilibrium implies a domain-conditional Great Unbundling: in high knowledge-velocity domains, value and employment shift toward many small specialized providers and thin orchestrators.

Key Points

• Coordination cost decomposition:
  • Total coordination cost C = Cintegration(E) + Cverification(T, {k_t}) + Cgovernance(T).
  • Cintegration is costs of maintaining integration edges; Cverification checks correctness and policy per task; Cgovernance is linear in throughput (audits/policy checks).
• Two coordination regimes:
  • Topology-dominated (SOA/microservices failure mode): E = Θ(n^2) and integration costs grow superlinearly with n.
  • Protocol-mediated hourglass: E = Θ(n) because providers integrate once to a shared protocol waist; integration costs are approximately linear in n.
• Sublinear Coordination Hypothesis (formal, testable): if per-task verification v(k) = O(k^γ) with γ < 1 (i.e., verification is reusable and amortizable across workflows), then total coordination grows sublinearly with system complexity and hourglass firms are efficient.
• Verification Reuse Condition (Lemma): a library of reusable outcome tests and policy templates implies vlocal(k)=O(k^γ), γ<1.
• Two falsifiable empirical predictions derived from the model:
• In a mature hourglass ecosystem, the marginal cost of adding an execution provider is approximately constant (i.e., near-linear scaling in n).
• The ratio of total coordination cost to task throughput remains stable as ecosystem size grows.
• Stability is conditional. Hourglass equilibrium holds if verification stays a throughput problem (reusable tests, outcome-based evaluation). It collapses toward re-centralization if verification or coupling becomes topology-dependent, protocols thicken into gatekeepers, liability centralizes, or SoR and platform effects reintroduce strong integration frictions.
• Macroeconomic implications: a domain-conditional “Great Unbundling” (shift of probability mass in firm-size distribution from large incumbents toward many micro-specialized agents) and a “missing middle” where the long tail of specialized providers gains aggregate share while very large firms can shrink headcount via automation.

Data & Methods

• Methodological approach: theoretical, minimal economic model + conceptual mapping to system design and empirical signatures.
  • Objects and notation: n providers, T tasks, each task t touches k_t providers; E = number of integration edges.
  • Cost decomposition into Cintegration(E), Cverification(T, {k_t}), Cgovernance(T).
  • Two structural cases for E: Θ(n^2) (pairwise integrations) vs Θ(n) (protocol waist).
  • Per-task verification split into vlocal(k) (reusable, local checks) and vcoupling(k) (cross-provider invariants); critical condition is vlocal(k) growing sublinearly.
• Formal results: lemmas and scaling relations that identify when hourglass architecture minimizes total coordination cost (requires γ < 1).
• Empirical testing suggestions (implicit in paper):
  • Measure onboarding/integration cost per new provider over time in candidate ecosystems (expect near-constant marginal cost in a mature hourglass).
  • Track total coordination-related expenditures (integration + verification + governance) per unit of throughput as ecosystem size grows (expect stability if hourglass holds).
  • Measure verification reuse: fraction of tests/policy artifacts reused across workflows, and empirical exponent γ via regression of verification cost vs workflow width k.
  • Compare sectors with varying knowledge velocity, SoR lock-in, and protocol standardizability to test the domain-conditional Great Unbundling.
• Data sources and priors referenced: empirical work on firm size distributions (Axtell), SOA/microservices literature (microservices premium), benchmarks and protocols in agent/tool research (ToolBench, MCP), and labor/automation literature (Autor; Acemoglu & Restrepo).
• Note: the paper is primarily conceptual/theoretical rather than reporting new empirical datasets; it presents falsifiable predictions and suggested empirical diagnostics.

Implications for AI Economics

• Firm boundaries and structure:
  • If the hourglass scaling holds, firms will unbundle: orchestration/intent interfaces and standardized protocols concentrate value at the waist while execution fragments into many vertical specialists (micro-agents).
  • The distribution of firm sizes flattens: more small specialist firms (long tail increases) and fewer intermediate “middle-sized” firms; large firms may shrink employee counts (automation efficiency) rather than disappear.
• Labor markets:
  • Rise of “scalable boutiques” and micro-specialized providers: demand for narrowly deep expertise that can be amortized across many clients.
  • Between-firm wage dispersion may increase (skill bundling and platform reputations); some superstar platform returns may persist at infrastructure level.
  • Roles in integration/verification/policy design (building reusable test harnesses, protocol design, observability for the waist) become high-value.
• Software and platform economics:
  • Protocol commoditization: successful standardized waists become infrastructural public goods that lower edge-to-edge integration cost; but they are also strategic chokepoints—risk of re-centralization if they become proprietary.
  • Value shifts upward (personalized GenUI/differentiated UX) and downward (deep vertical execution capability). Orchestrator incumbency is an economically strategic position if it controls the waist.
• Conditionality and risks:
  • The hourglass outcome is conditional on verification reuse and weak cross-provider coupling. If cross-provider invariants dominate (vcoupling large), or protocols consolidate into proprietary gatekeepers, the system re-centralizes.
  • Other constraining forces: platform concentration in model/cloud infrastructure, regulatory and capital requirements in some sectors, system-of-record (SoR) inertia, liability and compliance pressures.
• Policy and strategic takeaways:
  • Observers and regulators should monitor the evolution of protocol waists and verification ecosystems: openness, standardization, and reusability of tests/policies determine whether competition at execution layer is feasible.
  • Firms should invest in reusable verification artifacts, policy templates, and protocol-conformance tooling to participate in a headless ecosystem; alternatively, controlling thick-waist interfaces or SoR can be a defensible strategy to re-centralize.
• Empirical agenda: measure γ (verification growth exponent), marginal onboarding costs, and coordination-cost-to-throughput ratios across sectors to evaluate the paper’s falsifiable predictions and the extent of the Great Unbundling.

Conflict of interest note (from paper): authors are founders of Mantix, a company building infrastructure in this agentic-enterprise space.