The Agent Stack

The field has facts. It needs laws.

MAP surveyed how practitioners build agents. MAST cataloged how curated systems fail. HAL standardized evaluation. vLLM and Agentix made execution fast. None of them answers three questions: what the system is, why it behaves the way it does, and what the field should measure next. This project answers all three, with a layered reference model, a composition calculus, and a meta-measurement fitted on public data alone.

The page follows the argument in order. The ten open problems appear at the end because they are conclusions of the analysis, not premises of it.

The gap, why now, and what the entire literature missed

Objective. This project establishes the measurement science of AI agent systems: not whether agents can do tasks, but how well they work, how consistently, at what cost, and with what statistical rigor. The problem. The field can build and deploy agents, but it cannot yet characterize how they behave once deployed. No shared measurement methodology, reporting standard, or open trace corpus exists for agent systems.

Why now. Deployment is outpacing understanding, and the distance is measurable. Gartner projects that 40 percent of enterprise applications will embed task-specific AI agents by the end of 2026, up from less than 5 percent in 2025. On the other side of the ledger, MAP surveyed 306 practitioners across 26 domains and analyzed 86 deployed or pilot systems: 61.3 percent of deployed agents were never compared against any baseline, 74.2 percent are evaluated by manual human inspection, and practitioners rank reliability, scalability, latency, and cost as their single largest challenge.

What the whole literature missed, and why it matters. Benchmarks measure task accuracy but not workload structure. Reliability work, from pass^k to HAL, runs on benchmark logs, not production data. Serving research fabricates its workloads, because no public agent execution trace exists. Field studies report self-reported percentages, not distributions, tails, or time series. And nothing ties the pieces together. Measurement is the precondition for systems science: the Borg, Alibaba, and Azure trace releases each seeded a decade of research, and agents have no equivalent. This project builds the model, the laws, and the instruments that close the gap.

The canon: twenty-five works, surveyed and mapped

The literature review as cards instead of paragraphs. Each card opens to its core finding and the way this project uses it. Full citations live in the paper.

What the survey shows: four patterns

Reading the twenty-five works together, four regularities emerge. The chart below is computed directly from the cards above.

The reference model: eight layers, seven boundaries

The layers and boundaries are defined by this project; they are a contribution, not an inheritance. The model earns its keep by mapping: every surveyed work lands on a layer or a boundary. At three boundaries, prior work (MAST) has documented the failure modes. At the remaining four, the model predicts failure classes that no study has yet examined.

Select a layer to open its dossier: what is established, what is open, and the key works.

Coverage analysis: where the literature is dense and where it is empty

Each cell scores published coverage from 0 to 3, across five kinds of work. Select a cell. The empty cells are the point.

The composition calculus

Six laws govern how reliability, cost, and latency compose across steps, retries, verifiers, and agents. Five are implemented below from their closed forms; the tail law requires simulation. Each control sets a parameter of the law, and every readout follows from it.

Three controlled experiments, fitted on public data

No private dataset sits on the critical path. The laws are fitted on public artifacts, then completed by three experiments.

Public evidence base: MAST-Data HAL logs METR runs τ-bench SWE-bench WebArena GAIA BFCL AgentDojo AIOpsLab

Two instruments and the constants dashboard

The constants of agentic computing

The table the project exists to fill. Four rows are wired to the widgets above, so the dashboard moves as you explore; the rest await experiments E1 to E3.

Gap analysis across five research axes

The five axes are an organizing framework defined by this project; they emerge from where the twenty-five surveyed works cluster. For each axis, Established lists published results with their sources, and Open lists what no published work yet establishes: these are the targets this project sets for itself and for the community. Measurement science remains an axis of its own because it is the cross-cutting plane the other four consume.

Ten open problems, as conclusions

Each problem follows from the model, the coverage analysis, or a law. They close the argument; they do not open it.

Nine months with three students; the floor is thinking and writing

The three experiments are independent, so they parallelize perfectly across students. The model, the calculus, and the writing do not parallelize; they set the floor. Retrospective environment replay removes the calendar-time constraint from the drift experiment.

Budget for the nine-month default: about $420K. Minimum-viable at $180K: one student, experiment E1 only, laws fitted on existing public logs. Venues: SIGMETRICS/POMACS or SOSP/OSDI for the results; a CACM view piece for the agenda; community debut at the proposed AgentMetrics workshop, SIGMETRICS 2027. The AgentTrace schema ships with the measurement plane in month 3.

One anchor per layer

The stack dictates the guest list. This team is tentative: every entry below is a prospective collaborator to be invited, and no participation has been confirmed.

BibTeX

@misc{agentstack2026,
  title  = {The Agent Stack: A Reference Model, Composition Laws,
            and a Measurement Agenda for Agentic AI Systems},
  author = {Pasandi, Hannah B. and et al.},
  year   = {2026},
  note   = {Project page: https://your-domain.github.io/agent-stack}
}