AI 'agent teams' mirror organisational structure but lack human social glue; making context durable, inspectable and task-contingent — rather than copying human managerial rituals — produces better collective performance in simulations and LLM-agent traces.

Main Finding

Agentic AI behaves as a partial organizational analogue: multi‑agent systems reproduce many functional features of human organizations (differentiation of work, coordination, routines, boundary crossing, collective outcomes) but they are sustained not by social foundations (motivation, identity, trust, employment relations) and instead by context architecture (prompts, memory, traces, schemas, tools, validators, permissions). The central mechanism determining when agent collectives help or hurt is contextual transaction cost (CTC): the cost of making task context usable across agent and human boundaries. Empirical and simulation evidence shows that naive human‑imitation designs (many handoffs, free‑text debates, lossy summaries) often underperform, while agent‑native designs (shared state, durable/inspectable context, adaptive coordination) outperform when they reduce context loss, inspection/verification burdens, and correlated errors.

Key Points

• Partial analogue framing

  • Agent collectives are organizationally consequential (they differentiate work, coordinate interdependence, implement recurrent routines, cross boundaries, and generate collective knowledge/outcomes).
  • They are not social organizations: no socialization, identity, trust, moral accountability—behavior is enacted by contextual architecture (prompts, memory, validators, permissions).

• Contextual transaction cost (CTC)

  • CTC is the cost of moving and making task context usable across boundaries (handoffs, compression loss, semantic drift, verification burden, latency/compute, governance).
  • The paper formalizes event‑level CTCe (components: token/latency burden τ, handoff H, compression loss C, semantic drift D, verification burden V, governance G) and aggregates over traces to task/run level.
  • Collective efficiency CE combines quality and success probability with costs: CE = Q × P(success) / (1 + λ Cost). This emphasizes that small quality gains do not compensate for large contextual/verification costs.

• Mechanism implications (from theory and experiments)

  • Gains from agentic organization arise from specialization (Gs), parallelism (Gp), diversity (Gd), and verification (Gv). These must exceed added CTC, compute (K), and governance risk (Rg).
  • Human‑imitation forms (e.g., chains of roles mimicking human teams, free‑text debates) can add lossy handoffs, correlated deliberation, and heavy verification burdens—reducing collective efficiency.
  • Agent‑native forms (shared memory/state, explicit evidence attachments, adaptive coordination rules) perform better when they make context durable, inspectable, and task‑contingent.
  • Pseudo‑diversity risk: adding more agents does not guarantee independent information sources—correlated errors and manufactured consensus can eliminate the benefit of multiple agents unless diversity is engineered (different models, independent tools/evidence).

• Interface organizations

  • Effective human–agent collaboration requires explicit interface organizations: what evidence/traces must accompany agent outputs, mandatory human checkpoints, permissions for tools/actions, and disclosure of uncertainty.
  • Design of these interfaces determines whether agent outputs can be embedded in accountable human workflows.

Data & Methods

• Research strategy: computational theorizing combining synthetic simulations, instrumented real LLM agent traces, scaling experiments, and robustness checks. The objective is mechanism development rather than population estimation.

• Study 1 — Synthetic simulation

  • 8,000 synthetic knowledge‑work tasks.
  • Seven agent organization forms compared under identical tasks.
  • Controlled variation to identify task–organization interaction patterns (task‑fixed‑effects models, scaling of error correlation and communication modes).

• Study 2 — Real LLM traces

  • Trace‑instrumented runs with LLM‑based agent collectives.
  • Domains: software repair, long‑document question answering, legal review, literature synthesis.
  • Measured handoffs, memory writes/reads, tool calls, verifier activity, and verification burden to estimate CTC in actual executions.

• Study 3 — Robustness checks (reported in appendix)

  • Prompt sensitivity analyses, comparisons with open‑source models, and a prompt‑free reinforcement learning (DQN) organization selector to test whether observed patterns are robust to implementation choices.

• Metrics and analytic constructs

  • Event‑level CTC components (token/latency burden, handoffs, compression loss, semantic drift, verification burden, governance).
  • Aggregated run‑level CTC and a value decomposition: V = gains (specialization, parallelism, diversity, verification) − costs (CTC, compute, governance risk).
  • Collective efficiency CE = Q × P(success) / (1 + λ Cost), capturing tradeoffs between answer quality, reliability, and contextual/verification costs.

• Key empirical patterns reported

  • Human‑imitation multi‑agent workflows often underperform relative to compact or shared‑state architectures when they introduce lossy handoffs or require heavy verification.
  • Shared‑state and adaptive coordination architectures perform better when they preserve durable, inspectable context and reduce redundant deliberation.
  • Correlated error regimes and free‑text debates scale poorly; diversity must be structural (different evidence, models, tools) to yield real gains.

Implications for AI Economics

• Transaction cost economics extended

  • Contextual transaction costs reshape the tradeoffs determining optimal organizational form, outsourcing, and vertical integration for AI‑augmented tasks. Firms that internalize or reduce CTC (via shared memory, standards for evidence/traces, validators) can capture value that otherwise dissipates in verification and governance.

• Organization design and boundaries

  • The typical questions of make vs. buy now require assessing CTC (how hard is it to transfer usable context between agent providers and buyers?) rather than only compute or accuracy. High CTC favors internalizing agentic systems or investing in interface standards.

• Labor and task composition

  • Demand shifts toward roles that (a) design and maintain context architecture (prompts, memory, validators, schemas), (b) audit and interpret agent traces, and (c) adjudicate accountability. Routine tasks where CTCs are low are more easily automated; knowledge work with high context sensitivity retains human comparative advantage unless firms can lower CTC.

• Markets for agentic services and pricing

  • Pricing of agentic services should reflect not just compute and accuracy but also contextual support (evidence attachments, audit logs, explainability) because these lower downstream verification costs. Products that expose durable, inspectable context will command higher value in institutional settings (legal, medical, engineering).

• Returns to scale and competition

  • Winner‑take‑most dynamics are plausible: firms that develop superior interface organizations and durable shared state (reducing CTC) can scale agentic workflows more efficiently than firms that merely deploy many loosely coordinated agents. Conversely, poorly designed multi‑agent products that mimic human teams may scale poorly as handoffs and verification costs grow.

• Collective intelligence and market design

  • Simple aggregation markets or platforms that sell “more agents = better” will overpay unless they ensure independent evidence sources and low CTC. Market design should enforce standards for traceability, evidence preservation, and independence to realize gains from distributed agentic work.

• Governance, regulation, and policy

  • Regulators and standards bodies should focus on interface norms: required evidence/traces for automated outputs in accountable domains, minimum disclosure of uncertainty, audit logs, and tool/permission transparency. These interventions target CTC reduction and accountability alignment more directly than model accuracy mandates alone.

• Measurement and productivity accounting

  • Productivity measures must incorporate contextual and verification costs. Evaluating agentic automation by final output quality alone can be misleading; organizational economists and firms should measure CE‑style metrics that penalize expensive verification, slow turnarounds, and opaque traces.

Overall, the paper reframes the economics of AI‑enabled organizing: the key scarce asset is not only model accuracy or compute, but the capacity to make context durable, inspectable, and cheaply transferable across human and agent boundaries. Policies, firm strategies, and market designs that recognize and lower contextual transaction costs will determine where agentic AI creates net economic value.