Co-Tenant Interference Profiling

also known as Pre-Admission Interference Matrix, Sandbox Coexistence Profiling, Indirect Conflict Profiling

Before admitting a new autonomous controller to a fleet acting on one shared physical substrate, profile it in a sandbox against every incumbent, measure pairwise outcome degradation, and gate admission on the measured severity.

Context

A fleet of independently developed closed-loop controllers acts on one shared physical medium. The Near-RT RAN Intelligent Controller in an Open RAN deployment is the clearest instance: it hosts many xApps, each written by a different vendor, each an autonomous control loop over radio access network parameters, each admitted to the platform on its own merits. The same shape appears wherever separately authored agents actuate one substrate — warehouse robots on one floor, bidding agents in one auction, heating and shading controllers in one building. Each controller is reviewed alone, declares which parameters it writes, and is judged safe on that declaration.

Problem

Controllers that share a substrate rarely collide on a parameter; they collide through the medium that couples the parameters they each own separately. One xApp owns transmit power, another owns resource-block allocation or handover offset. They never write the same field, so a static reading of their action spaces shows no overlap and the design review finds nothing. Radio physics couples them anyway: both move cell coverage, and the damage shows up only as degraded outcomes. PACIFISTA measures it — a 16% throughput loss between xApps with similar goals, and up to 30% performance degradation plus instability between xApps with opposing goals. Worse, the coupling is conditional rather than constant: two xApps may conflict under certain conditions and coexist under others, so a single review, or a single test run, cannot settle whether the pair is safe.

Forces

A declared action space is cheap to check and easy to reason about, but it describes which fields a controller writes, not which outcomes it moves, and indirect coupling lives entirely in the gap between the two.
Each controller is individually correct and individually approved, so there is no faulty component to find; the harm exists only in the pair, and no single-agent review can see it.
Conflicts are context-dependent — a pair that coexists under light load can oscillate under mobility or congestion — so evidence gathered under one operating condition does not transfer to another.
Profiling every candidate against every incumbent across a sweep of conditions grows with the square of the fleet, and sandbox time is the scarce resource that decides how thorough the gate can be.
The actuation changes physical conditions for thousands of subscribers at once, so discovering the interaction from production telemetry means the degradation has already been served to users, which is why the measurement has to happen before admission rather than as a live rollback.

When to use

Several independently developed controllers actuate one shared physical or economic medium and are admitted to the fleet one at a time.
Controllers write disjoint parameters, so a comparison of their action spaces cannot show a conflict that the medium nonetheless creates.
The damage from a conflicting pair is emergent over sustained joint operation rather than visible in any single action.
A sandbox exists that reproduces the substrate faithfully enough for outcome measurements to transfer to production.
Actuation reaches many users at once, so discovering the conflict from production telemetry means the degradation has already been served.

When not to use

Controllers contend directly on the same field or resource, where locking, a single-writer path or a priority rule settles it without profiling.
No sandbox reproduces the coupling mechanism, in which case a clean profile is misleading and a runtime mediator or staged rollout is the honest remedy.
The fleet is small, stable and authored by one team that can reason about the coupling from the design directly.
Actions are cheap to reverse and low-blast-radius, where observing a live canary costs less than a full pairwise sweep.
Controllers change faster than the profiling sweep can complete, so every matrix is stale before it is used.

Example

An office building runs two separate automation controllers: one raises and lowers the window blinds to manage glare, the other sets the air conditioning. Neither can write the other's setting, so a review of what each is allowed to change finds no overlap and both are switched on. On a sunny afternoon the blind controller opens for daylight, the cooling controller fights the heat that comes in with it, and the two cycle against each other all day. A week of paired sandbox runs across different weather conditions would have shown the swing before either went near the real building.

Diagram

flowchart TD A["Candidate controller requests admission"] --> B{"Declared action spaces overlap?"} B -- "no overlap on paper" --> C["Pair candidate with each incumbent"] C --> D["Sandbox: solo A, solo B, joint run"] D --> E["Sweep operating conditions: load, mobility, traffic mix"] E --> F["Measure outcome deltas against solo baselines"] F --> G{"Pairwise degradation severity"} G -- "none measured" --> H["Admit; record interference matrix"] G -- "mild" --> I["Admit with narrowed actuation scope or a mediator"] G -- "severe or unstable" --> J["Refuse; report the pair and the condition"] B -- "overlap on paper" --> K["Settle directly: locking, single writer, priority rule"] H --> L["Any fleet or substrate change invalidates the matrix"] I --> L L --> C

Solution

Therefore:

Stand up a sandbox that reproduces the substrate faithfully enough for outcomes to be comparable — an emulated radio environment, a plant digital twin, a physics simulation of the floor. Keep a registry of the incumbent controllers currently admitted to the fleet. When a candidate arrives, run three configurations for every incumbent and every declared operating condition: the candidate alone, the incumbent alone, and the two together. Collect end outcomes rather than actions — throughput, tail latency, handover failures, oscillation amplitude — and compare the joint run against each solo baseline with a statistical model, so that run-to-run noise is separated from real degradation. The result is an interference matrix over the fleet: one severity score per pair per condition, plus the conditions under which the pair is safe. The admission gate reads that matrix. A candidate with no measurable degradation joins. A candidate with mild degradation joins with its actuation scope narrowed, or with a mediator placed between it and the substrate for the parameters implicated. A candidate that drives an incumbent into instability is refused, and the refusal names the pair and the condition rather than a general objection. Because the matrix is a property of the fleet, admitting anyone invalidates it: re-profile when a controller is added, updated, or when the substrate itself changes.

What it gives you

Conflicts that no static comparison of action spaces can reveal become measurable, because the measurement is taken in outcome space where the coupling actually shows.
The gate produces a specific artefact — this pair, under this condition, costs this much throughput — so a refusal can be argued with and a restriction can be scoped to the parameter implicated.
Degradation is discovered before subscribers or occupants experience it, which matters when one admission changes physical conditions for thousands of users at once.
The interference matrix accumulates into a map of which controllers can share the substrate, useful long after the admission decision that produced each entry.

What it costs you

Cost grows with the square of the fleet multiplied by the condition sweep, so late admissions get either a long queue for sandbox time or thinner coverage than early ones.
Pairwise profiling misses three-way interactions; a candidate can be clean against every incumbent individually and still destabilise the fleet once all of them run together.
The sandbox is only as good as its fidelity to the substrate — an emulator that smooths over the physical coupling reports a clean pair and admits the conflict it was built to catch.
The matrix goes stale the moment any co-tenant is updated, and a fleet under continuous delivery can spend more time re-profiling than operating.
Severity thresholds are a policy choice with no natural value; set them tight and useful controllers are refused, set them loose and measured degradation is admitted on purpose.

What this pattern forbids. A controller cannot be admitted to a fleet sharing one substrate on the strength of a declared non-overlapping action space; it must first be profiled in the sandbox against every incumbent across the declared operating conditions, and admission is refused, or its actuation scope narrowed, when measured pairwise degradation exceeds the fleet's threshold. The resulting interference matrix is only valid for the fleet composition that produced it, so no controller may be added or updated without re-profiling.

And the patterns that stand alongside it, or against it —

complementsSimulate Before Actuate★— Before issuing an irreversible action, run a deterministic simulation that computes pre-conditions, invariants, and expected deltas; require a verifier — automated or human — to green-light the simulated outcome before the real command is sent.
complementsShadow Canary★★— Run a candidate agent version in shadow alongside the champion, comparing outputs without affecting users.
complementsRace Conditions on Shared Tool Resources✕— Anti-pattern: let concurrent agents perform read-modify-write on shared external resources without locking, producing silent data corruption.
complementsPriority Matrix (Conflict Resolution)★— Pre-define how the agent must resolve specific classes of goal conflicts via a human-authored lookup table — transforming the agent from a decision-maker (where it fails on competing objectives) into a decision-implementer.
complementsControl-Loop-Mapped Agent Chain·— Give each existing feedback loop in a plant's control chain its own operator agent carrying that loop's control-theoretic context, and resolve their competing proposals with the chain's own deterministic selector logic.
complementsStigmergic Coordination★★— Agents coordinate indirectly by leaving and reading marks in a shared environment (files, queues, scratchpads, world model) so that one agent's trace stimulates another's next action, with no direct messaging.
complementsAgent Capability Manifest★— Let each agent publish a standardized self-description — identity, skills, endpoint, and auth needs — at a well-known location, so others discover it and bind by capability at runtime instead of through hardcoded coupling.
complementsRed-Team Sandbox Reproduction★— Routinely re-reproduce canonical alignment-failure modes inside a sealed sandbox per release; treat the alignment regression suite as a deployment gate.

Neighbourhood

Click any neighbour to follow the language. Scroll to zoom, drag to pan.

References

Provenance

Source: patterns/co-tenant-interference-profiling.md on GitHub · commit c389d0e · view history
Added to catalog: 2026-09-10
Last updated: 2026-09-10
Contribute: open an issue or PR at github.com/agentpatternscatalog/patterns.