\documentclass[11pt,a4paper]{article} % ---- Packages ---- \usepackage[utf8]{inputenc} \usepackage[T1]{fontenc} \usepackage{amsmath,amssymb} \usepackage{graphicx} \usepackage{booktabs} \usepackage{hyperref} \usepackage{xcolor} \usepackage{listings} \usepackage{subcaption} \usepackage{tikz} \usetikzlibrary{shapes,arrows.meta,positioning,fit,calc,matrix} \usepackage[numbers]{natbib} \usepackage{geometry} \usepackage{enumitem} \geometry{margin=1in} % ---- Colors ---- \definecolor{highfit}{HTML}{2E7D32} \definecolor{medfit}{HTML}{F57F17} \definecolor{lowfit}{HTML}{C62828} \definecolor{neutral}{HTML}{546E7A} % ---- Title ---- \title{% From Factory Floor to Token Stream:\\ A Taxonomy of Operations Management Methods\\ for LLM Agent Orchestration% } \author{ Christian Nennemann\\ Independent Researcher\\ \texttt{chris@nennemann.de} } \date{April 2026} \begin{document} \maketitle % ============================================================ \begin{abstract} Multi-agent systems built on large language models (LLMs) increasingly adopt metaphors from human project management---sprints, standups, code review---yet draw from a remarkably narrow slice of the operations management literature. This paper presents a systematic taxonomy of twelve established PM/OM methods, evaluates their structural compatibility with LLM agent constraints (stateless invocation, cheap cloning, deterministic dysfunction, absence of human psychology), and identifies which methods are underexploited, which are inapplicable, and which require fundamental adaptation. We find that methods designed for \emph{flow optimization} (Kanban, Theory of Constraints) and \emph{rapid decision-making} (OODA Loop) are structurally well-suited to agent orchestration but remain largely unexplored, while methods centered on \emph{human psychology} (Scrum ceremonies, Design Thinking empathy phases) transfer poorly without significant reformulation. We propose a decision framework for selecting orchestration methods based on task complexity, agent count, and quality requirements, and identify five open research directions at the intersection of operations management and agentic AI. \end{abstract} % ============================================================ \section{Introduction} \label{sec:intro} The dominant paradigm for multi-agent LLM systems borrows from agile software development: agents are organized into ``teams'' with role-based specialization, tasks are decomposed into work items, and results are reviewed before merging \citep{hong2024metagpt, qian2024chatdev}. This borrowing is natural---the humans building these systems are software engineers familiar with agile methods---but it is also narrow. The operations management literature contains dozens of methods developed over a century of industrial practice, each encoding different assumptions about workflow structure, quality assurance, failure modes, and coordination costs. Not all of these methods are equally applicable to LLM agents. Agents differ from human workers in five structurally important ways: \begin{enumerate}[label=\textbf{C\arabic*}] \item \label{c:stateless} \textbf{Stateless invocation}: Agents do not retain memory between invocations unless explicitly persisted. Human team members accumulate institutional knowledge automatically. \item \label{c:cloning} \textbf{Cheap to clone, expensive to coordinate}: Spawning a new agent costs milliseconds and cents; coordinating two agents costs tokens and latency. For human teams, the inverse holds---hiring is expensive, coordination is (comparatively) cheap. \item \label{c:dysfunction} \textbf{Deterministic dysfunction}: LLM agents fail in predictable, repeatable patterns---verbosity, scope creep, false positives---rather than the varied, context-dependent failures of human cognition \citep{nennemann2026archeflow}. \item \label{c:psychology} \textbf{No psychology}: Agents have no morale, fatigue, ego, or office politics. Methods designed to manage human psychology (retrospectives, team-building, conflict resolution) have no direct function. \item \label{c:speed} \textbf{Cycle speed}: Agents complete tasks in seconds to minutes, enabling iteration frequencies that would be impractical for human teams. Methods that assume week-long or month-long cycles can be compressed. \end{enumerate} These constraints define a \emph{fitness landscape}: some PM/OM methods gain effectiveness when applied to agents (because agents remove friction those methods were designed to manage), while others lose their raison d'\^etre (because they solve human problems agents don't have). This paper contributes: \begin{itemize} \item A systematic taxonomy of twelve PM/OM methods evaluated against the five agent constraints (\ref{c:stateless}--\ref{c:speed}). \item A compatibility matrix scoring each method's structural fit for agent orchestration (\S\ref{sec:matrix}). \item A decision framework for practitioners selecting orchestration strategies (\S\ref{sec:decision}). \item Five open research directions at the intersection of operations management theory and agentic AI (\S\ref{sec:future}). \end{itemize} % ============================================================ \section{Background: Current Agent Orchestration Landscape} \label{sec:background} \subsection{Frameworks and Their Implicit PM Models} The current generation of multi-agent LLM frameworks implicitly adopts project management concepts, though rarely with explicit attribution to PM/OM theory. \textbf{MetaGPT} \citep{hong2024metagpt} assigns human job titles (product manager, architect, engineer) and enforces communication through Standardized Operating Procedures (SOPs)---an implicit adoption of \emph{waterfall} phase gates with role-based access control. \textbf{ChatDev} \citep{qian2024chatdev} simulates a software company with sequential phases (design, coding, testing, documentation). Despite the ``company'' framing, the execution model is a \emph{linear pipeline} with pair-programming-style chat between adjacent roles. \textbf{AgileCoder} \citep{nguyen2024agilecoder} is the first framework to explicitly adopt sprint-based iteration, assigning Scrum Master and Product Manager roles to LLM agents with a Dynamic Code Graph Generator tracking inter-file dependencies between sprints. \textbf{CrewAI} organizes agents into ``crews'' with a ``manager'' agent orchestrating task delegation---an implicit \emph{hierarchical management} model with single-point-of-failure coordination. \textbf{AutoGen} \citep{wu2023autogen} provides a conversation-based framework where agents negotiate through multi-turn dialogue. The implicit model is \emph{committee decision-making}---all agents see all messages, consensus emerges through discussion. \textbf{The Six Sigma Agent} \citep{patel2026sixsigma} decomposes tasks into atomic dependency trees, executes each node $n$ times with independent LLM samples, and uses consensus voting to achieve defect rates scaling as $O(p^{\lceil n/2 \rceil})$---reaching 3.4 DPMO (the Six Sigma threshold) at $n=13$. \textbf{Reflexion} \citep{shinn2023reflexion} implements a de facto PDCA loop through verbal reinforcement: Plan $\to$ Act $\to$ Evaluate (Check) $\to$ Reflect (Act), though it does not name this structure explicitly. \textbf{ArcheFlow} \citep{nennemann2026archeflow} explicitly applies PDCA quality cycles with Jungian archetypal roles, representing the first framework to deliberately adopt a named PM/OM methodology with formal convergence criteria. \subsection{The Gap} Despite the variety of frameworks, the PM/OM methods actually employed cluster tightly around four approaches: (1) waterfall-style sequential phases (MetaGPT, ChatDev), (2) role-based team simulation (CAMEL \citep{li2023camel}, CrewAI), (3) informal ``manager'' delegation (AutoGen), and (4) agile sprints (AgileCoder). The Six Sigma Agent \citep{patel2026sixsigma} is a notable exception---the only framework to explicitly name a PM/OM method as its primary architectural contribution. Methods from lean manufacturing, constraint theory, military decision-making, innovation management, and failure analysis remain unexplored in the peer-reviewed agent orchestration literature, despite strong structural compatibility with agent constraints. % ============================================================ \section{Taxonomy of PM/OM Methods} \label{sec:taxonomy} We evaluate twelve methods spanning five categories: iterative improvement, flow optimization, decision-making, innovation management, and quality engineering. For each method, we describe the core mechanism, evaluate structural compatibility with agent constraints \ref{c:stateless}--\ref{c:speed}, identify the primary adaptation required, and assess overall fitness. % ---- 3.1 Iterative Improvement ---- \subsection{Iterative Improvement Methods} \subsubsection{PDCA (Plan--Do--Check--Act)} \label{sec:pdca} \textbf{Origin}: Shewhart \citep{shewhart1939statistical}, popularized by Deming \citep{deming1986out}. \textbf{Mechanism}: Four-phase cycle repeated until quality targets are met. Each cycle narrows the gap between current and desired state through structured feedback. \textbf{Agent fitness}: \textsc{High}. PDCA's phase structure maps directly to agent orchestration: Plan (research + design agents), Do (implementation agent), Check (review agents), Act (routing + merge decisions). The cycle abstraction handles the core challenge of ``when to stop iterating'' through convergence metrics. Demonstrated in ArcheFlow \citep{nennemann2026archeflow}. \textbf{Key adaptation}: Convergence detection must be automated (human PDCA relies on subjective judgment). ArcheFlow addresses this with a convergence score based on finding classification (new, resolved, persistent, regressed) and oscillation detection. \textbf{Constraint fit}: Stateless (\ref{c:stateless})---artifacts persist state between cycles. Cloning (\ref{c:cloning})---fresh agents per cycle avoid accumulated bias. Speed (\ref{c:speed})---cycles complete in minutes, enabling 2--3 cycles where humans would manage one. \subsubsection{Scrum} \label{sec:scrum} \textbf{Origin}: Schwaber \& Sutherland, 1995. \textbf{Mechanism}: Time-boxed sprints with defined roles (Product Owner, Scrum Master, Development Team), ceremonies (planning, daily standup, review, retrospective), and artifacts (backlog, sprint board, burndown). \textbf{Agent fitness}: \textsc{Low--Medium}. Scrum's ceremony-heavy structure exists primarily to manage human coordination challenges: standups maintain shared awareness (agents can share a filesystem), retrospectives address interpersonal friction (agents have none), sprint planning negotiates capacity (agents have deterministic throughput). The useful kernel---time-boxed work with a prioritized backlog---is trivially implementable without Scrum's overhead. \textbf{Key adaptation}: Strip ceremonies, keep the backlog + sprint structure. ``Daily standups'' become status file reads. ``Retrospectives'' become cross-run memory extraction. The Scrum Master role is pure overhead for agents. \textbf{Constraint fit}: Psychology (\ref{c:psychology})---most Scrum ceremonies solve human problems. Speed (\ref{c:speed})---sprint length compresses from weeks to minutes. Cloning (\ref{c:cloning})---team stability (a Scrum value) is irrelevant when agents are stateless. \subsubsection{DMAIC (Six Sigma)} \label{sec:dmaic} \textbf{Origin}: Motorola, 1986; systematized by General Electric. \textbf{Mechanism}: Define--Measure--Analyze--Improve--Control. Unlike PDCA, DMAIC emphasizes \emph{statistical measurement} of process capability and explicitly separates analysis (understanding the problem) from improvement (fixing it). \textbf{Agent fitness}: \textsc{Medium--High}. The Define--Measure--Analyze front-loading is valuable for agents: it forces explicit quality metrics \emph{before} implementation, preventing the common failure mode of agents optimizing for the wrong objective. The Control phase---establishing monitoring to prevent regression---maps to cross-run memory systems. \textbf{Key adaptation}: Agents can compute statistical process control metrics (defect rates, cycle times, sigma levels) automatically from event logs. The ``Measure'' phase, which is expensive and tedious for humans, becomes a strength: agents can instrument everything. \textbf{Constraint fit}: Speed (\ref{c:speed})---full DMAIC in minutes. Dysfunction (\ref{c:dysfunction})---agent failure modes have measurable baselines, making sigma calculations meaningful. Stateless (\ref{c:stateless})---Control phase requires persistent monitoring, which must be explicitly built. % ---- 3.2 Flow Optimization ---- \subsection{Flow Optimization Methods} \subsubsection{Kanban} \label{sec:kanban} \textbf{Origin}: Toyota Production System, Taiichi Ohno, 1950s. \textbf{Mechanism}: Pull-based workflow with explicit work-in-progress (WIP) limits. Work items flow through columns (stages); new work is pulled only when capacity is available. No iterations---continuous flow. \textbf{Agent fitness}: \textsc{High}. Kanban's WIP limits directly address a critical agent challenge: \emph{coordination cost scaling}. Without WIP limits, spawning more agents increases throughput initially but eventually degrades quality due to coordination overhead (conflicting changes, merge conflicts, context fragmentation). Kanban provides a principled mechanism for determining optimal concurrency. \textbf{Key adaptation}: WIP limits should be \emph{dynamic}, adjusting based on observed coordination costs (merge conflicts, finding duplications) rather than fixed. The pull mechanism maps naturally: agents poll a task queue and pull the highest-priority item they can handle. \textbf{Constraint fit}: Cloning (\ref{c:cloning})---WIP limits are \emph{exactly} the missing constraint for cheap-to-clone agents. Speed (\ref{c:speed})---flow metrics (lead time, cycle time, throughput) update in real-time. Psychology (\ref{c:psychology})---no ``swarming'' or ``blocked item'' social dynamics to manage. \subsubsection{Theory of Constraints (TOC)} \label{sec:toc} \textbf{Origin}: Goldratt, \emph{The Goal}, 1984. \textbf{Mechanism}: Identify the system's constraint (bottleneck), exploit it (maximize its throughput), subordinate everything else to it, elevate it (invest to remove it), repeat. The Five Focusing Steps. \textbf{Agent fitness}: \textsc{High}. In multi-agent pipelines, the bottleneck is typically the most capable (and expensive) agent: the implementation agent that must run on a powerful model, or the security reviewer that requires deep context. TOC provides a framework for organizing the entire pipeline around this constraint. \textbf{Key adaptation}: ``Exploit the constraint'' means ensuring the bottleneck agent never waits for input. Pre-compute its context, batch its inputs, and schedule cheaper agents (research, formatting, validation) to run during its processing time. ``Subordinate'' means cheaper agents should produce output in the format the bottleneck needs, not in whatever format is easiest for them. \textbf{Constraint fit}: Cloning (\ref{c:cloning})---non-bottleneck agents are cheap to overprovision. Speed (\ref{c:speed})---constraint shifts can be detected and responded to within a single run. Dysfunction (\ref{c:dysfunction})---bottleneck agent's failure mode has outsized impact, justifying targeted shadow detection. \subsubsection{Lean / Toyota Production System} \label{sec:lean} \textbf{Origin}: Ohno, 1988; Womack \& Jones, 1996. \textbf{Mechanism}: Eliminate waste (\emph{muda}), reduce variability (\emph{mura}), avoid overburden (\emph{muri}). Seven wastes: overproduction, waiting, transport, overprocessing, inventory, motion, defects. \textbf{Agent fitness}: \textsc{Medium--High}. The seven wastes map surprisingly well to agent systems: \begin{itemize}[nosep] \item \textbf{Overproduction}: Agents generating output nobody reads (verbose research reports, unused alternative proposals). \item \textbf{Waiting}: Agents idle while waiting for predecessor output (sequential pipeline where parallel would work). \item \textbf{Transport}: Redundant context passing (sending full codebase to agents that need only a diff). \item \textbf{Overprocessing}: Running thorough review on trivial changes. \item \textbf{Inventory}: Accumulated artifacts from prior cycles that are never referenced. \item \textbf{Motion}: Agents reading files they don't need, exploring irrelevant code paths. \item \textbf{Defects}: Findings that are false positives, requiring rework to dismiss. \end{itemize} \textbf{Key adaptation}: Lean's ``respect for people'' pillar has no direct analog. The technical pillar (continuous improvement, waste elimination) transfers fully. % ---- 3.3 Decision-Making ---- \subsection{Decision-Making Methods} \subsubsection{OODA Loop (Observe--Orient--Decide--Act)} \label{sec:ooda} \textbf{Origin}: John Boyd, 1976. Military strategy for air combat; later generalized to competitive decision-making. \textbf{Mechanism}: Continuous loop of Observe (gather data), Orient (analyze context, update mental models), Decide (select course of action), Act (execute). The key insight is that the \emph{speed} of the loop---not any individual decision's quality---determines competitive advantage. ``Getting inside the opponent's OODA loop'' means acting faster than the adversary can react. \textbf{Agent fitness}: \textsc{High}. OODA is structurally similar to PDCA but optimized for speed over thoroughness. For agent systems, this maps to scenarios requiring rapid adaptation: adversarial testing, incident response, market-reactive coding, or any context where the problem space changes during execution. \textbf{Key adaptation}: Boyd's ``Orient'' phase---updating mental models based on new information---is the hardest to implement for stateless agents. It requires either persistent state (a world model that updates across iterations) or a ``fast reorientation'' agent that rapidly synthesizes new information into an updated context. \textbf{Constraint fit}: Speed (\ref{c:speed})---agents can OODA at superhuman frequency. Stateless (\ref{c:stateless})---the Orient phase needs explicit state management. Psychology (\ref{c:psychology})---Boyd's concept of ``mental agility'' translates to model selection: smaller, faster models for rapid OODA; larger models for deep Orient phases. \subsubsection{Cynefin Framework} \label{sec:cynefin} \textbf{Origin}: Snowden \& Boone, 2007. \textbf{Mechanism}: Classify problems into five domains---\textsc{Clear} (obvious cause-effect), \textsc{Complicated} (expert analysis needed), \textsc{Complex} (emergent, probe-sense-respond), \textsc{Chaotic} (act first, then sense), \textsc{Confused} (unknown domain)---and apply domain-appropriate strategies. \textbf{Agent fitness}: \textsc{Medium--High}. Cynefin provides a \emph{meta-framework}: instead of choosing one orchestration method for all tasks, classify the task first, then select the appropriate method: \begin{itemize}[nosep] \item \textsc{Clear}: Single agent, no review (``fix this typo''). \item \textsc{Complicated}: Expert agent with review (PDCA fast workflow). \item \textsc{Complex}: Multiple competing proposals, let results emerge (PDCA standard/thorough with parallel alternatives). \item \textsc{Chaotic}: Act immediately, stabilize, then analyze (OODA with hotfix agent, then PDCA for proper fix). \end{itemize} \textbf{Key adaptation}: Task classification must be automated. Proxies: number of files affected, cross-module dependencies, security sensitivity, test coverage of affected area. % ---- 3.4 Innovation Management ---- \subsection{Innovation Management Methods} \subsubsection{Stage-Gate} \label{sec:stagegate} \textbf{Origin}: Cooper, 1990. \textbf{Mechanism}: Innovation projects pass through stages (scoping, business case, development, testing, launch), separated by gates where a cross-functional team decides: Go, Kill, Hold, or Recycle. The gate decision is binary---no ``continue with reservations.'' \textbf{Agent fitness}: \textsc{Medium}. The gate mechanism maps well to agent confidence checks: a Creator agent's proposal either meets the confidence threshold (Go) or doesn't (Kill/Recycle). However, Stage-Gate assumes expensive stages (weeks/months of human work), making Kill decisions high-stakes. For agents, stages are cheap (minutes), reducing the value of formal gate decisions. \textbf{Key adaptation}: Gates become lightweight confidence checks rather than committee reviews. The ``Kill'' decision---rare and painful in human innovation---should be common and cheap for agents. Explore multiple proposals in parallel, gate aggressively, continue only the best. \subsubsection{Design Thinking} \label{sec:designthinking} \textbf{Origin}: IDEO / Stanford d.school, 2000s. \textbf{Mechanism}: Five phases: Empathize (understand the user), Define (frame the problem), Ideate (generate solutions), Prototype (build quickly), Test (get feedback). Emphasis on user empathy and divergent thinking. \textbf{Agent fitness}: \textsc{Low}. Design Thinking's core value proposition---\emph{empathy with users}---is precisely what LLM agents cannot genuinely do. Agents can simulate empathy (generate persona-based scenarios), but the insight that comes from observing real users in context has no agent equivalent. The Ideate phase (divergent brainstorming) is feasible but produces quantity over quality without the ``empathy filter'' that makes Design Thinking effective. \textbf{Key adaptation}: If used, the Empathize phase must be replaced with explicit user research artifacts (personas, journey maps, interview transcripts) provided as input. This transforms Design Thinking from a discovery method into a synthesis method---fundamentally changing its nature. \subsubsection{TRIZ} \label{sec:triz} \textbf{Origin}: Altshuller, 1946--1985. Theory of Inventive Problem Solving. \textbf{Mechanism}: Problems contain contradictions (improving one parameter worsens another). TRIZ provides a contradiction matrix mapping 39 engineering parameters to 40 inventive principles. Instead of compromise, TRIZ seeks solutions that resolve the contradiction. \textbf{Agent fitness}: \textsc{Medium}. TRIZ's structured problem-solving is well-suited to agents: the contradiction matrix is a lookup table, and agents can systematically apply inventive principles. However, TRIZ requires \emph{reformulating the problem as a contradiction}---a creative step that is itself challenging for agents. \textbf{Key adaptation}: Provide the contradiction matrix as context. Train agents to identify the ``improving parameter'' and ``worsening parameter'' in engineering tasks (e.g., ``improving security worsens performance''). Use TRIZ principles as a structured brainstorming prompt for the Creator archetype. % ---- 3.5 Quality Engineering ---- \subsection{Quality Engineering Methods} \subsubsection{FMEA (Failure Mode and Effects Analysis)} \label{sec:fmea} \textbf{Origin}: US Military, 1949; adopted by automotive (AIAG) and aerospace. \textbf{Mechanism}: For each component/process step, systematically enumerate: (1) potential failure modes, (2) effects of each failure, (3) causes, (4) current controls, (5) risk priority number (severity $\times$ occurrence $\times$ detection). Address highest-RPN items first. \textbf{Agent fitness}: \textsc{High}. FMEA's systematic enumeration is exactly what LLM agents excel at: given a design, enumerate everything that could go wrong, assess severity, and propose mitigations. The Risk Priority Number provides a quantitative framework for prioritizing review effort---more principled than the common ``CRITICAL/WARNING/INFO'' severity classification. \textbf{Key adaptation}: Use FMEA \emph{before} implementation (as part of the Plan phase) rather than only during review. An FMEA agent analyzes the Creator's proposal and generates a failure mode table; the Maker then implements with awareness of high-RPN failure modes; the Guardian validates that mitigations are in place. \textbf{Constraint fit}: Dysfunction (\ref{c:dysfunction})---agents' own failure modes can be pre-enumerated via FMEA, creating a meta-level quality system. Cloning (\ref{c:cloning})---FMEA agents are cheap (analytical, not creative), enabling systematic coverage. \subsubsection{Statistical Process Control (SPC)} \label{sec:spc} \textbf{Origin}: Shewhart, 1920s. \textbf{Mechanism}: Monitor process outputs over time using control charts. Distinguish \emph{common cause} variation (inherent to the process) from \emph{special cause} variation (attributable to specific events). React only to special causes; reduce common cause variation through process improvement. \textbf{Agent fitness}: \textsc{Medium--High}. SPC requires historical data, which agent orchestration systems naturally generate (event logs, finding counts, cycle times, token usage). Control charts over agent effectiveness scores can distinguish between normal variation (``Guardian found 2 issues this run vs. 1 last run'') and genuine degradation (``Guardian's false positive rate spiked after a model update''). \textbf{Key adaptation}: Sufficient run history is needed to establish control limits. Early runs operate without SPC; after 10--20 runs, control limits become meaningful. Model updates reset control limits (new process = new baseline). % ============================================================ \section{Compatibility Matrix} \label{sec:matrix} Table~\ref{tab:matrix} scores each method against the five agent constraints, producing an overall fitness assessment. \begin{table}[t] \centering \small \caption{Compatibility matrix: PM/OM methods scored against agent constraints. \textcolor{highfit}{\textbf{+}} = method benefits from this constraint; \textcolor{lowfit}{\textbf{--}} = method is undermined; \textcolor{neutral}{\textbf{0}} = neutral. Overall fitness: H = High, M = Medium, L = Low.} \label{tab:matrix} \begin{tabular}{@{}l*{5}{c}c@{}} \toprule \textbf{Method} & \textbf{C1} & \textbf{C2} & \textbf{C3} & \textbf{C4} & \textbf{C5} & \textbf{Fit} \\ \midrule PDCA & \textcolor{neutral}{0} & \textcolor{highfit}{+} & \textcolor{highfit}{+} & \textcolor{neutral}{0} & \textcolor{highfit}{+} & \textbf{H} \\ Scrum & \textcolor{lowfit}{--} & \textcolor{neutral}{0} & \textcolor{neutral}{0} & \textcolor{lowfit}{--} & \textcolor{highfit}{+} & \textbf{L--M} \\ DMAIC & \textcolor{lowfit}{--} & \textcolor{highfit}{+} & \textcolor{highfit}{+} & \textcolor{neutral}{0} & \textcolor{highfit}{+} & \textbf{M--H} \\ Kanban & \textcolor{neutral}{0} & \textcolor{highfit}{+} & \textcolor{neutral}{0} & \textcolor{highfit}{+} & \textcolor{highfit}{+} & \textbf{H} \\ TOC & \textcolor{neutral}{0} & \textcolor{highfit}{+} & \textcolor{highfit}{+} & \textcolor{highfit}{+} & \textcolor{highfit}{+} & \textbf{H} \\ Lean & \textcolor{neutral}{0} & \textcolor{highfit}{+} & \textcolor{neutral}{0} & \textcolor{lowfit}{--} & \textcolor{highfit}{+} & \textbf{M--H} \\ OODA & \textcolor{lowfit}{--} & \textcolor{highfit}{+} & \textcolor{neutral}{0} & \textcolor{highfit}{+} & \textcolor{highfit}{+} & \textbf{H} \\ Cynefin & \textcolor{neutral}{0} & \textcolor{neutral}{0} & \textcolor{neutral}{0} & \textcolor{neutral}{0} & \textcolor{neutral}{0} & \textbf{M--H} \\ Stage-Gate & \textcolor{neutral}{0} & \textcolor{highfit}{+} & \textcolor{neutral}{0} & \textcolor{highfit}{+} & \textcolor{lowfit}{--} & \textbf{M} \\ Design Think. & \textcolor{neutral}{0} & \textcolor{neutral}{0} & \textcolor{neutral}{0} & \textcolor{lowfit}{--} & \textcolor{neutral}{0} & \textbf{L} \\ TRIZ & \textcolor{neutral}{0} & \textcolor{highfit}{+} & \textcolor{neutral}{0} & \textcolor{neutral}{0} & \textcolor{highfit}{+} & \textbf{M} \\ FMEA & \textcolor{neutral}{0} & \textcolor{highfit}{+} & \textcolor{highfit}{+} & \textcolor{highfit}{+} & \textcolor{highfit}{+} & \textbf{H} \\ SPC & \textcolor{lowfit}{--} & \textcolor{neutral}{0} & \textcolor{highfit}{+} & \textcolor{highfit}{+} & \textcolor{highfit}{+} & \textbf{M--H} \\ \bottomrule \end{tabular} \end{table} \subsection{Analysis} Several patterns emerge from the compatibility matrix: \textbf{High-fitness methods share three properties}: they are \emph{mechanistic} (decisions follow rules, not judgment), \emph{flow-oriented} (optimize throughput, not team dynamics), and \emph{metric-driven} (quality is quantified, not discussed). PDCA, Kanban, TOC, OODA, and FMEA all share this profile. \textbf{Low-fitness methods are psychology-dependent}: Scrum and Design Thinking derive their primary value from managing human cognitive and social limitations. Without those limitations, the methods become overhead. \textbf{The ``Cheap Clone'' constraint is universally beneficial}: every method either benefits from or is neutral to the ability to spawn agents cheaply. This suggests that agent orchestration should generally favor \emph{parallelism}---run multiple approaches simultaneously, then select the best result. \textbf{``Stateless'' is the most disruptive constraint}: methods that assume accumulated knowledge (Scrum's team velocity, SPC's control charts, DMAIC's baseline measurements) require explicit persistence mechanisms that agents don't provide natively. % ============================================================ \section{Hybrid Approaches and Method Composition} \label{sec:hybrid} The methods in our taxonomy are not mutually exclusive. Effective agent orchestration likely requires combining methods at different levels: \subsection{Proposed Three-Layer Architecture} \begin{description} \item[Strategic layer (Cynefin)]: Classify the task and select the appropriate orchestration method. Simple tasks get a single agent; complicated tasks get PDCA; complex tasks get parallel competing approaches; chaotic tasks get OODA. \item[Operational layer (PDCA/OODA + Kanban)]: Execute the selected method with flow control. Kanban WIP limits prevent coordination overload. PDCA provides quality convergence for standard tasks; OODA provides rapid adaptation for time-sensitive tasks. \item[Quality layer (FMEA + SPC + TOC)]: Monitor execution quality. FMEA front-loads failure analysis in the Plan phase. SPC monitors long-term agent effectiveness trends. TOC identifies and optimizes around bottleneck agents. \end{description} \subsection{ArcheFlow as a Case Study} ArcheFlow \citep{nennemann2026archeflow} already implements elements of this three-layer architecture, though without explicitly naming all methods: \begin{itemize}[nosep] \item \textbf{Strategic}: Workflow selection (fast/standard/thorough) functions as a simplified Cynefin classification. \item \textbf{Operational}: PDCA cycles with convergence detection; sprint mode with WIP-limited parallel dispatch (implicit Kanban). \item \textbf{Quality}: Shadow detection (behavioral FMEA for agent failure modes); effectiveness scoring (rudimentary SPC); Guardian fast-path (TOC---don't waste the bottleneck on clean code). \end{itemize} The gap is in explicit TOC application (identifying and optimizing around the most expensive agent) and in OODA integration for time-sensitive tasks. % ============================================================ \section{Decision Framework} \label{sec:decision} We propose a practitioner-oriented decision framework for selecting orchestration methods based on three dimensions: \begin{figure}[h] \centering \begin{tikzpicture}[ box/.style={draw, rounded corners, minimum width=3.5cm, minimum height=0.7cm, font=\small, fill=#1}, arrow/.style={-{Stealth[length=3mm]}, thick}, ] % Decision tree \node[box=yellow!20] (start) {Task arrives}; \node[box=orange!15, below=0.8cm of start] (cynefin) {Classify (Cynefin)}; \node[box=green!15, below left=1cm and 2cm of cynefin] (clear) {Clear}; \node[box=green!15, below left=1cm and 0cm of cynefin] (complicated) {Complicated}; \node[box=blue!10, below right=1cm and 0cm of cynefin] (complex) {Complex}; \node[box=red!10, below right=1cm and 2cm of cynefin] (chaotic) {Chaotic}; \node[box=white, below=0.7cm of clear, text width=2.5cm, align=center, font=\scriptsize] (m1) {Single agent\\No review}; \node[box=white, below=0.7cm of complicated, text width=2.5cm, align=center, font=\scriptsize] (m2) {PDCA fast\\+ FMEA}; \node[box=white, below=0.7cm of complex, text width=2.5cm, align=center, font=\scriptsize] (m3) {PDCA thorough\\+ parallel proposals}; \node[box=white, below=0.7cm of chaotic, text width=2.5cm, align=center, font=\scriptsize] (m4) {OODA\\then PDCA}; \draw[arrow] (start) -- (cynefin); \draw[arrow] (cynefin) -- (clear); \draw[arrow] (cynefin) -- (complicated); \draw[arrow] (cynefin) -- (complex); \draw[arrow] (cynefin) -- (chaotic); \draw[arrow] (clear) -- (m1); \draw[arrow] (complicated) -- (m2); \draw[arrow] (complex) -- (m3); \draw[arrow] (chaotic) -- (m4); \end{tikzpicture} \caption{Decision framework for selecting agent orchestration method based on Cynefin task classification.} \label{fig:decision} \end{figure} \textbf{Cross-cutting concerns} apply regardless of classification: \begin{itemize}[nosep] \item \textbf{Kanban WIP limits}: Always. Prevents coordination overload. \item \textbf{TOC awareness}: Identify the costliest agent; schedule others around it. \item \textbf{SPC monitoring}: After 10+ runs, establish control limits for agent effectiveness. \item \textbf{Lean waste audit}: Periodically review token usage patterns for waste (unused artifacts, redundant context, overprocessing). \end{itemize} % ============================================================ \section{Open Research Directions} \label{sec:future} \subsection{Adaptive Method Selection} Current frameworks use a fixed orchestration method. An adaptive system would classify each incoming task (Cynefin), select the appropriate method, and switch methods mid-execution if the task's nature changes (e.g., a ``complicated'' task reveals unexpected complexity during exploration). This requires a \emph{method-aware orchestrator} that understands the assumptions and exit criteria of each method. \subsection{Kanban for Agent Swarms} As agent counts increase beyond 5--10, coordination costs dominate. Kanban's WIP limits and flow metrics provide a theoretical basis for determining optimal agent concurrency, but empirical studies are needed to establish how coordination cost scales with agent count across different task types and model capabilities. \subsection{OODA for Adversarial Agent Scenarios} Boyd's OODA loop was designed for competitive environments where speed of decision-making determines the winner. Applications include adversarial testing (red team agents vs. blue team agents), competitive code generation (multiple agents racing to solve a problem), and incident response (rapid diagnosis and mitigation under time pressure). \subsection{Cross-Method Quality Metrics} Each PM/OM method defines quality differently: PDCA uses convergence scores, Six Sigma uses sigma levels, Lean uses waste ratios, SPC uses control limits. A unified quality metric for agent orchestration---one that allows meaningful comparison across methods---does not yet exist. \subsection{FMEA for Agent Failure Modes} Agent failure modes (hallucination, scope creep, false positive reviews, persona drift \citep{lu2026assistant}) can be systematically enumerated using FMEA methodology. A comprehensive FMEA catalog for LLM agents---with severity, occurrence, and detection ratings calibrated from empirical data---would provide a foundation for designing more robust orchestration systems. % ============================================================ \section{Conclusion} \label{sec:conclusion} The operations management literature offers a rich toolkit for agent orchestration that extends far beyond the agile methods currently dominant in the field. Our taxonomy reveals that the highest-fitness methods---PDCA, Kanban, TOC, OODA, and FMEA---share a common profile: mechanistic, flow-oriented, and metric-driven. Methods centered on human psychology (Scrum, Design Thinking) transfer poorly without fundamental reformulation. The key insight is that LLM agents are not ``fast humans.'' They have fundamentally different constraint profiles---cheap to clone, expensive to coordinate, stateless, psychologically inert---and these differences make some PM/OM methods \emph{more} effective (OODA loops at superhuman speed, FMEA with exhaustive enumeration) while rendering others irrelevant (standups without psychology, retrospectives without learning). We encourage the agent orchestration community to look beyond agile sprints and role-playing frameworks toward the broader operations management tradition. A century of industrial practice has much to teach us about orchestrating intelligent agents---if we take the time to translate. % ============================================================ \section*{Acknowledgments} The author thanks the operations management and quality engineering communities whose work, developed over decades for human organizations, provides the theoretical foundation for this analysis. % ============================================================ \bibliographystyle{plainnat} \bibliography{taxonomy-refs} \end{document}