Literature Review

Coordination cost in heterogeneous AI tool ecosystems

Working draft · v0.1 · 2026-04-29 · For posting at peopleanalyst.com/research/devplane

This review situates the DevPlane research program within five literatures whose intersection defines the program's contribution. Each section: what the field knows, the canonical references, and where the C1 lead study (see PROPOSAL.md) extends or differs.

A note on confidence. References marked [C] are canonical and high-confidence — I am confident they exist as cited and represent the field accurately. References marked [F] are field-level claims (consensus or pattern across the literature) where I have not verified a specific source. Open questions marked [DR] are flagged for OpenAI Deep Research expansion (see deep-research-prompts.md). Mike will validate uncertain references against scite, Google Scholar, or the deep-research outputs before public posting.

---

1. Cockpit and process-control HCI — the automation tradition

This literature is the program's primary anchor. It studies what happens when humans supervise automation that does work formerly done by humans, in high-stakes settings (aviation cockpits, nuclear control rooms, anesthesia, autonomous vehicles).

The foundational paper is Bainbridge's Ironies of Automation [C], a six-page article in Automatica in 1983 that articulated two ironies: (1) designer errors are absorbed and amplified by automation, and (2) automating the easy parts of a task leaves the operator with the harder residual, while simultaneously eroding the practice and situation awareness needed to handle that residual. Forty years later it remains the most-cited starting point for human-factors work on automation.

The literature elaborates Bainbridge along several axes. Endsley's situation awareness model [C] (1995, in Human Factors) provides the standard three-level decomposition: perception, comprehension, projection. SAGAT, Endsley's measurement instrument, has been used in hundreds of automation studies. Lee & See (2004) [C], also in Human Factors, gives the field its definitive treatment of trust calibration: trust is not just "more is better" but a calibration problem with both over-trust and under-trust costs, mediated by the operator's mental model of automation reliability. Parasuraman & Manzey (2010) [C] consolidate the literature on complacency and automation bias as attentional phenomena — operators systematically under-allocate attention to checks the automation purportedly handles.

The risk-compensation strand sits adjacent to but distinct from this literature. Peltzman (1975) [C], in the Journal of Political Economy, observed that automotive safety regulation produced smaller-than-expected reductions in fatalities, attributing the gap to compensating increases in risky driving. Wilde's risk homeostasis theory [C] generalized this into a hypothesis that operators target a constant level of perceived risk, fully offsetting safety improvements. The empirical record supports partial offset more often than full homeostasis [F], but the partial-offset prediction is the live one.

Recent work on automation in software development is sparse [DR]. The cockpit literature has been applied to clinical decision support, autonomous vehicles, and aviation, but the AI-agents-on-a-codebase setting is not well represented in the canonical literature as of this writing. This is one of the gaps the C1 study targets directly.

Where DevPlane research extends. The cockpit literature was built around a narrow operator-supervises-one-or-two-systems pattern. Heterogeneous AI tool ecosystems present a different topology: one operator supervises N concurrent agents with overlapping authority over the same codebase, mediated through a coordination surface (the kanban/registry). The Ironies prediction in this regime has not been formally tested. C1 tests it.

2. Computer-supported cooperative work — the coordination tradition

CSCW has spent four decades studying how groups coordinate work through shared artifacts, and is the natural literature for stigmergic coordination patterns of the kind DevPlane uses.

Schmidt & Bannon's articulation work [C] (1992, in the inaugural issue of the Computer Supported Cooperative Work journal) is the field's foundational concept: the ongoing meta-work of dividing, scheduling, and integrating cooperative work, distinct from the productive work itself. Articulation work is exactly what DevPlane's coordination layer is trying to absorb.

Suchman's Plans and Situated Actions [C] (1987) reframed workplace activity from plan-execution to situated improvisation. The implication for coordination tools is sharp: coordination layers that treat plans as authoritative and execution as derivative will systematically misrepresent how work actually happens. DevPlane's design tension between assignment-as-spec and assignment-as-orientation echoes this directly.

Grudin's "Why CSCW applications fail" [C] (1988, Communications of the ACM) named the social asymmetries that doom most groupware: the people who do the coordination work are not the people who get the benefit. Translated to DevPlane: the operator absorbs the cost of writing dispatches; the agents (and their downstream consumers) get the benefit. This asymmetry is a design constraint, not an incidental fact.

The stigmergy strand within coordination research draws from biology (Theraulaz on ant trail formation) [F] and AI (Hayes-Roth's blackboard architectures from the 1980s) [C]. Stigmergic coordination — actors coordinating through modifications to a shared environment rather than through direct messages — is the theoretical category DevPlane occupies. The CSCW community has revisited stigmergy periodically [DR], but a focused contemporary literature on stigmergic coordination among AI agents is thin.

Recent CSCW work on human-AI collaboration has grown rapidly since 2020 [DR]. CHI and CSCW conference proceedings include a growing stream of empirical studies on developers using single AI assistants (Copilot, ChatGPT). The multi-agent operational case is less well covered.

Where DevPlane research extends. Three contributions: (1) empirical telemetry on stigmergic coordination at sustained operational scale, where most stigmergy work has been theoretical or simulation-based; (2) a contemporary instance of Grudin's coordination-cost asymmetry in the AI-agent setting; (3) a setting where Suchman's situated-action framing can be tested against the formal-plan framing using continuous data on dispatch-text fidelity to shipped output.

3. Empirical software engineering — the methodology tradition

The empirical SE literature provides the methodological template for studying coordination using repository telemetry.

Brooks' The Mythical Man-Month [C] (1975) made the original argument that adding people to a software project increases coordination cost super-linearly. Conway's Law [C] (1968) named the structural relationship between organizational communication patterns and the systems they produce. Both are pre-empirical observations whose empirical testing has been the field's project for decades.

Herbsleb & Mockus [C] (in IEEE Transactions on Software Engineering) on coordination costs in distributed software development, particularly comparing same-site vs. distributed work. These are large-N observational studies using version-control and issue-tracker data, and their methodological approach — operationalize coordination cost as time-to-completion deltas under specified conditions — is directly applicable to the C1 study at smaller scale with richer per-event data.

The contemporary empirical SE literature uses GitHub and similar repositories at very large N. Vasilescu and collaborators at CMU [DR] have published extensively on developer activity, productivity measurement, and the methodological challenges of telemetry-based research (selection effects, what counts as a "unit," the ecological validity of public-repo data). Their methodological discipline — particularly around acknowledging confounds and treating activity metrics with appropriate skepticism — is what the DevPlane research program is trying to inherit at smaller N.

A specific recent strand is the empirical study of GitHub Copilot and similar AI assistants [DR]. RCTs from Microsoft and academic groups have measured task-completion time, satisfaction, and code quality. These typically study individual developers using a single AI assistant on bounded tasks, not multi-agent operational settings.

Where DevPlane research extends. The empirical SE literature has rich methodology and large-N observational corpora, but very small operational telemetry on the multi-agent operator role. DevPlane provides exactly that: small N, very high resolution. The methodological commitments of empirical SE (acknowledged confounds, skeptical treatment of activity metrics) are what protect the program from the failure modes of less-disciplined "AI productivity" research.

4. Behavioral decision-making — the operator tradition

The B-arm of the program (human-AI interaction) draws principally from this literature, and even the C1 study depends on it for the trust-calibration mechanism.

Kahneman & Tversky's heuristics-and-biases program [C] supplies the broad framing: human judgment under uncertainty is systematically biased in directions predictable from a small number of mechanisms (representativeness, availability, anchoring). Tetlock's work on forecasting calibration [C] (Expert Political Judgment, 2005; Superforecasting, 2015 with Gardner) provides the canonical methodology for measuring calibration: Brier scores, calibration diagrams, the distinction between calibration and resolution.

Klein's naturalistic decision-making [C] and the recognition-primed decision model offer a complementary framing for expert decision-making in time-pressured operational settings, which is exactly the operator-of-multiple-agents situation.

Decision fatigue [C] (Vohs, Baumeister, and others, though this literature has had replication concerns [F]) predicts that decision quality degrades across a session of repeated choices. The competing prediction from deliberate practice literature (Ericsson) [C] is that performance improves within a session as the operator warms up. These literatures make opposite predictions on the dispatch-quality-across-session question (B-arm secondary), which makes that question worth asking — the data adjudicates.

The trust-in-automation literature discussed in §1 is the bridge from behavioral decision-making into the cockpit/HCI tradition: Lee & See specifically integrate trust theory from Mayer/Davis/Schoorman with the automation context.

Where DevPlane research extends. The behavioral literature has rich theory and laboratory evidence, but very little continuous data on a single operator making thousands of micro-decisions over months in an operational setting. DevPlane's dispatch corpus is unusually well-suited to test calibration, fatigue, and continuation-bias predictions that the laboratory literature can only approximate.

5. Multi-agent systems and stigmergic coordination — the architecture tradition

The A-arm of the program draws on a literature that overlaps CSCW (§2) but extends into AI and distributed systems.

Hayes-Roth's blackboard architectures [C] (1985, in the Artificial Intelligence journal) defined the model of multiple specialist agents coordinating through a shared structured artifact rather than through direct messaging. Stigmergic coordination in DevPlane is a direct contemporary instance.

Distributed-systems coordination primitives [C] — consensus protocols, supervision trees, work queues — provide the engineering vocabulary for what DevPlane does at the agent-coordination layer. The CSCW vs. distributed-systems literatures have historically not communicated well [F]; DevPlane operates at exactly their intersection.

Recent work on multi-agent LLM systems [DR] — agentic workflows, agent orchestration frameworks (LangGraph, CrewAI, AutoGen, others), multi-agent debate/critique architectures — is rapidly growing but largely engineering-focused rather than empirical. The empirical question of how multi-agent LLM coordination actually fails in production is under-studied.

Where DevPlane research extends. A1's stigmergic-drift question — does coordination through shared artifacts produce systematically different failure signatures than direct-communication coordination — is a question this literature implies but rarely tests with real production data on heterogeneous agents.

Synthesis — where this program sits

The DevPlane research program is at the intersection of these five literatures, and its contribution is partly to put them in conversation with each other:

• From the automation tradition, we take the Ironies-of-Automation prediction and the trust-calibration mechanism (foundation for C1)
• From CSCW, we take articulation-work, situated-action, and the coordination-asymmetry observations (foundation for the A-arm)
• From empirical SE, we take the methodological discipline for telemetry-based observation (foundation for measurement)
• From behavioral decision-making, we take the calibration measurement instruments (foundation for B-arm)
• From multi-agent systems, we take the stigmergic-coordination architectural vocabulary (frame for A1)

The program's distinctive position is the continuous production telemetry on a real heterogeneous-AI multi-agent operator role. Each of the five literatures has rich theory and partial empirics; none has had access to this kind of data at this resolution.

Open questions for deep-research expansion

These are the gaps where I have low confidence and where targeted deep research is highest-value. They map directly to prompts in deep-research-prompts.md.

1. Post-2020 application of Ironies of Automation to AI/LLM contexts. The prediction is specific; how often has it been formally tested in any AI-related setting, and with what results?
2. Recent CMU HCII / S3D / LTI work on human-AI collaboration in software development. Specific authors, recent papers (2023–2026), live citations to the cockpit-automation literature.
3. Risk compensation outside automotive. How robust is the partial-offset finding across non-automotive domains? What's the most rigorous extension to information-work or knowledge-work settings?
4. Empirical studies of multi-agent LLM coordination failures. What's been published, what corpora are available, what failure-mode taxonomies already exist?
5. Stigmergic coordination, contemporary computational treatment. Has the CSCW or AI literature returned to stigmergy in the last decade? With what theoretical extensions?

The deep-research outputs will be reconciled against this review and the review revised before posting.

Concluding note

A literature review's job is to make clear what is known, what is contested, and what is unknown. By that standard, the AI-coordination-cost question sits at the intersection of much that is known (cockpit automation, CSCW coordination, empirical SE methodology), some that is contested (decision fatigue replication, full-vs-partial risk homeostasis), and a substantial amount that is genuinely unknown — particularly the joint behavior of heterogeneous AI agents and a single human operator in a sustained operational setting. The DevPlane research program is sized to address the unknowns in this last category, with as much methodological discipline as the existing traditions provide.