AgenticStack.
network systems and measurement for deployable agents
a research agenda led by Hannah B. Pasandi · Scientist, UC Berkeley Sky Computing Lab
in plain terms

An AI agent finishes one task by making many model calls behind the scenes, but the networks, schedulers, and cost accounting underneath were all built for a single call at a time. AgenticStack first measures what agent workloads actually do on the network, then builds the serving, reliability, and safety layers that make them dependable in production. Each project below is one piece of that stack.

the agenda

Agentic AI raised the unit of computation from the call to the plan. The networks, schedulers, and accounting beneath agents still see single calls. AgenticStack raises the systems and the measurement to match.

AgenticStack is a network-systems and measurement agenda for deployable agents: it measures what agent workloads do on the wire, then builds the control plane that makes them reliable. Measurement first, mechanism second. Each project below is one primitive in that stack. The pattern behind them, in the spirit of Specula, is that every AI-generated or AI-consumed signal in production needs a systems contract: a fidelity contract for generated code, a reliability contract for verifier signals, a telemetry contract for traces, and a safety contract for control actions.

expose autonomy as policy, enforce trust as mechanism.
the area

Network systems and measurement for deployable agents

Work on agents and networks tends to sit in one of two places. One treats the network as the medium between agents, an ecosystem for autonomous agents to discover and coordinate; this is largely a 6G and wireless vision (AgentNet, SANNet, agentic cores). The other puts agents in charge of running the network: agentic NetOps. This agenda is a third thing, and a less crowded one: the network-systems and measurement view of the agent workloads themselves.

It treats those workloads as the object of study: how to serve and schedule plans rather than calls, and what contracts the data plane needs to keep plan-level autonomy reliable. The tradition is measurement-driven systems work in the style of SIGCOMM, NSDI, SoCC, and SOSP, not a vision paper. The landscape at the end places this lane against its neighbors.

the shift

One workflow, two views

An operator who sees only model calls cannot account for an agent. The same workflow looks like noise at the call level and like a budgeted unit at the plan level. AgenticStack makes the plan the visible object. Numbers are illustrative.

GET /v1/messages  gen     312 tok   req a91f
GET /v1/messages  gen     880 tok   req a920
GET /v1/messages  verify  140 tok   req a921
GET /v1/messages  retry   880 tok   req a922
GET /v1/messages  verify  140 tok   req a923
GET /v1/messages  judge   205 tok   req a924
GET /v1/messages  gen     410 tok   req a925  (abandoned step)
no plan id . no workflow budget . abandoned work is invisible _
plan 7c2e  "summarize and file the contract"
  stage generate     x2     1192 tok
  stage verify       x2      280 tok    reliability 0.86, defer if < 0.80
  stage retry        x1      880 tok
  stage judge        x1      205 tok
  stage abandoned    x1      410 tok    charged to plan, not hidden
  -----------------------------------------
  workflow total 2967 tok . budget 3500 . reliability 0.86 . status COMPLETE _
illustrative trace, not measured results
architecture

Four layers, one visible unit

The agent runtime stays free to plan and act. The control plane is where trust is enforced: it owns the plan ID, the workflow budget, the reliability interval, the certificate, and the event log, and it sits over the serving and network data plane. Operators read outcomes, not calls.

Agent runtime plans · reasoning traces · tool calls · retries · verify and judge free to plan and act AgenticStack control plane plan ID · workflow budget · reliability interval · certificate · event log trust enforced here: accounting, budgets, contracts, gates, rollback Serving & network data plane gateway · router · edge / device / satellite tiers · verifier tier · protocol deployment moves the work under latency, energy, carbon, quota, availability Operator-visible outcomes reliable completion · safety · cost · latency · privacy · carbon
Fig 1 - the AgenticStack control plane between an agent runtime and the serving / network data plane.
primitives

Each paper is one primitive in the stack

The agenda is not a list of unrelated systems. Each project supplies one contract or mechanism the control plane needs.

accounting

Plan-level accounting

Charge generator, verifier, retry, judge, and abandoned steps to the plan that caused them, not to individual calls.

→ Single-Call Fallacy
budget

Reliability budget

A scheduler that trades single-request latency for workflow reliability under a bounded budget.

→ Patient Packets
prefetch

Plan-aware prefetch

Read the agent's plan while generation is in flight and stage state before the next step starts.

→ PALSE
contract

Reliability contract

The router consumes a reliability vector, a verdict interval, and a defer action instead of trusting a verdict as truth.

license

Deployment license

An untrusted proposer plus a small certificate that decides what may run; agents never control unsafe systems directly.

→ Vouch · MAC-Agent · Loom
telemetry

Telemetry & reasoning tax

Expose the boundary between hidden reasoning tokens and user-visible tokens so schedulers can see test-time compute.

→ TraceProbe · Reasoning Tax
privacy gate

Action-level privacy gate

Contextual-integrity gates on the agent's actions: tool calls, memory, paralinguistic inference.

→ Voice-agent privacy papers
serving & systems

Inference runtime and OS-level work

System / titleVenueWhat it is
PALSE: Plan-Aware Prefetch for Agentic LLM Serving at the Edge
submitted
IEEE/ACM SEC '26Edge serving middleware that reads the agent's reasoning trace as a prefetch oracle: plan reader, plan-conditioned prefetch, speculative coherence. Reports 90.5% SLO vs 74.1% for the strongest reactive baseline at 12 concurrent agents.
SCF: Single-Call Fallacy (OS framing)
submission
AgenticOS @ SOSP '26Plan-level OS accounting for agents on MAST traces. Per-call accounting is blind to plan structure; the OS should expose the plan, not the call.
SkyCascade
in planning
MobiArchDevice-edge-cloud speculative agent execution; cross-tier speculative agents for mobile LLM serving.
measurement & reliability

What deployed agents actually do, measured

System / titleVenueWhat it is
Confine
submitted
HotNets '26MAST annotator-agreement measurement plus a reliability-contract routing API (per-mode reliability vector, verdict interval, defer action) for LLM cascades. Project page →
Wire / Three LLMs
in planning
SIGMETRICS '27 (Fall R2)Packet-level measurement separating chat vs inference-serving vs agentic traffic. Three workloads, different on every network metric examined.
Gauge
in submission
In submissionFailure and safety measurement of deployed agents.
Reasoning Tax + TraceProbe
accepted
HotCarbon '26The hidden reasoning trace carries 87.1% of output tokens and 78.4% of operational carbon per agent task; a trace-length budget removes 56 to 71% of carbon at a 2.3-point accuracy cost. TraceProbe, a 90-line vLLM patch, exposes the trace-to-answer split in per-call telemetry at under 0.2% overhead.
networked & radio-edge agents

Agents over real links, and protocol synthesis

System / titleVenueWhat it is
Patient Packets
in planning
SoCC '26 (R2)Reliability-budgeted scheduler at the agent gateway: spends an application-declared byte budget for verification, retry, and judging with three rules and a (1+ε) bound; trades single-request latency for workflow reliability. Project page →
Old Models / VarRouter
in planning
MLSys '27Multi-version LLM coexistence plus a version-aware routing primitive.
MAC-Agent
submission
IEEE ICNPSafe MAC-protocol synthesis with an R0 epidemic-threshold safety certificate. Agentic modules BLOCKMINE (spec extraction) and DRIFTREASON (bounded drift diagnosis) around a masked tabular-RL core.
Loom
in submission
In submissionLLM agent for MAC-layer control.
Cardinal
in planning
SIGCOMMLLM-driven MAC protocol synthesis under an R0 stability certificate. The dense-MAC depth anchor of the protocol-synthesis line.
Heddle
program, 3 papers
SIGCOMM lineAgentic protocol-synthesis framework spanning Wi-Fi, 5G, NTN; a multi-paper program with an open block-level testbed.
OrbitMesh
in planning
Four-tier device / cellular / cloud / direct-to-cell-satellite mobile LLM serving.
privacy & safety of agents

Threat models, leakage, and defenses

System / titleVenueWhat it is
Beyond the Signal: Privacy in Agentic Voice Assistants
submission
SPSC '26 (short)Frames the four leakage channels that open once a voice assistant becomes an agent that plans, calls tools, and remembers across speakers.
Acoustic Context Injection
submission
SPSC '26 (long)Attack plus the CI-Gate defense built around the agent's actions: tool calls, memory, paralinguistic inference. Systems version aimed next at PoPETs '27 and USENIX Security '27.
proposed, not yet written

Ideas with a target

IdeaVenueWhat it is
the landscape

Where this work sits

The sections above are the work itself. The landmarks below place this network-systems and measurement lane against its neighbors, the 6G agent-to-agent vision and the agentic-NetOps line, so the boundary is clear. The page is meant to be read and cited as an entry point to the area.

Toward Agentic AI Networking in 6G: A Generative Foundation Model-as-Agent Approach (AgentNet)
IEEE Communications Magazine, 2025
The 6G vision of a networking ecosystem for agents to discover, coordinate, and share knowledge. Network as the medium between agents.
SANNet: A Semantic-Aware Agentic AI Networking Framework for Multi-Agent Cross-Layer Coordination
2025
Studies agentic systems from the communication-networking perspective; notes most agent work ignores the networking architecture.
The Agentic-AI Core (A-Core): an AI-empowered core network for next-generation telecom
Engineering / ICC Workshops, 2025
A mission-oriented agentic core network. The telecom-infrastructure framing.
Workshop @ SOSP 2026
OS primitives, isolation, scheduling, and observability for agent workloads. The OS-side neighbor.
Workshop @ SIGCOMM 2026
Networks for training clusters: topology, routing, synchronization. Adjacent, but training-side rather than agent-workload-side.
The Future of NetOps Is Agentic
Gartner, 2026 (industry)
Agents that operate the network. The NetOps framing, distinct from measuring and serving agent workloads.
cite this page
Hannah B. Pasandi. "AgenticStack: Network Systems and Measurement for Deployable Agents." hanabhp.github.io/agentic_ai.html, 2026.