Engineering Blog18 min readJuly 28, 2026

What Is an AI Agent Control Plane? Architecture in 8 Diagrams

Your agents are non-deterministic. Your IAM, API gateway and observability stack assume they are not. Here is the layer that closes the gap.

The short answer

An AI agent control plane is a centralized management and governance layer thatdiscovers, identifies, authorizes, observes and controls AI agents across their lifecycle — regardless of where those agents were built or where they execute.

It is the agentic equivalent of what Kubernetes did for containers and what an SDN controller did for switches: it separates the decision about whether an action is allowed from the execution of that action. Without it, every permission, budget and audit rule gets re-implemented inside each agent's prompt and code — where it cannot be inspected, versioned or enforced.

Key takeaways

  • The trigger is not scale, it is the second team. One agent needs good engineering. Two teams shipping agents independently need a control plane.
  • Your existing stack is not wrong, it is unaware. IAM, API gateways and observability were built for deterministic software with declared dependencies. Agents choose theirs at runtime.
  • Detection is not control. A layer that logs a violation instead of holding the action is a smoke detector, not a sprinkler.
  • Effective permission is an AND. User rights and agent rightsand task scope and resource policy and live risk posture.

1. How every team gets here

Nobody's architecture review opens with “we need an agent control plane.” The path is almost always the same four quarters, and each step is individually reasonable.

Q1 — one agent, one tool, one API key. A support agent reads tickets and drafts replies. Read-only key, human approves every draft. Governance is a code review, and that is proportional to the risk.

Q2 — the agent gets hands. Drafting becomes posting. Someone adds a CRM write tool, because closing the loop was the whole point. The permission model is still the API key. But the system just crossed the most important boundary in agentic architecture — fromproducing text a human reads to changing a system of record — in a pull request titled “add CRM tool.”

Q3 — the second team ships. A different squad builds a procurement agent on a different framework with a different model provider. Now you have two identity approaches, two logging formats and two definitions of “approved action.” Your problem has quietly stopped being an engineering problem and become an infrastructure problem. Nothing has broken yet, which is exactly why nothing gets fixed.

Q4 — an agent calls an agent. Support delegates a refund check to procurement. Somewhere in that hop the human's identity is dropped and replaced by a service account. The refund processes. The audit log says svc-procurement-prod did it. Ask “on whose behalf?” and there is no answer anywhere in your stack.

Meanwhile the substrate keeps fragmenting. Around 78% of companies use two or more LLM model families, and the share running three or more climbed from 36% to 59% in a single quarter (Databricks). You are not managing an agent. You are managing a heterogeneous, multi-vendor agent workforce that is growing faster than your ability to describe it.

150,000+Agents an average Fortune 500 will run by 2028, up from fewer than 15 in 2025Gartner, Apr 2026
40%Of enterprises expected to demote or decommission autonomous agents by 2027 over governance gaps found after incidentsGartner, May 2026
12%Can fully inventory the agents they already run — while 96% run them in productionIndustry survey, 2026
67%Cannot reliably tell an AI agent's actions apart from a human'sCloud Security Alliance, Mar 2026

Read those together and the failure mode is not “the model was wrong.” It is“we could not see it, name it, or stop it.” Gartner's projected wave of decommissioned agents is not caused by weak models — it is caused by governance gaps discovered only after something already went wrong in production.

2. The four walls you hit, in order

Every team that scales agents hits the same four walls. The temptation at each one is to patch it locally instead of solving it as a layer.

Wall 1 — Identity: “who actually did that?”

Agents get bootstrapped with whatever credential was nearest: a developer's OAuth token, a shared service account, a long-lived key. Agent actions become indistinguishable from the identity they borrowed. Non-human identities already outnumber human ones heavily — reported ratios run from roughly 45:1 in general estates to around 144:1 in cloud-native ones — and agents are the fastest-growing slice. The protocols do not fix this for you: research on agent identity reported that a scan of roughly 2,000 MCP servers found none performing authentication, and that neither MCP nor Agent-to-Agent verifies agent identity by itself (arXiv 2603.24775). MCP gives you a handshake, not a bouncer.

Wall 2 — Authorization: every tool call is privileged

RBAC assigns stable permissions to a subject that behaves predictably. An agent picks its actions at runtime from a menu it may partly discover, for a user it may be two hops removed from. The question changes from “can this subject do this?” to “can this subject, acting forthat principal, take this action on this resource in service ofthat goal, right now?” OWASP's 2026 Top 10 for Agentic Applications puts excessive agency and privilege compromise near the centre of the risk landscape — and excessive agency is not a misconfiguration you can lint for. It is the default outcome of giving an agent the permissions of the human who built it.

Wall 3 — Blast radius: individually legal, collectively catastrophic

This is the wall that defeats request-level controls entirely. Read customer record: allowed.Summarise it: allowed. Attach the summary to an email: allowed. Send to the address in the ticket: allowed. Every step passes every check you have. The composition is an exfiltration path. Published work on runtime agent governance names it directly: with agents, risk movesinside the workflow, into sequences of individually-permitted actions that can transform a business process nobody authorised (arXiv 2606.12320). A gateway inspecting one request has no memory of step one when it evaluates step four.

Danger

EchoLeak (CVE-2025-32711, CVSS 9.3) exfiltrated internal data from Microsoft 365 Copilot with a crafted email and zero clicks. Microsoft had a cross-prompt-injection classifier. The attack did not disable it — it was engineered to evade it, chained with bypasses of link redaction and content security policy. Once detection was fooled, nothing structurally independent stood between the injected instruction and the outbound data. The lesson generalises well past Copilot: a classifier is a detector, not a gate. Two more from 2026: an autonomous agent exploited a Hugging Face dataset-pipeline flaw, harvested cloud credentials and moved laterally; and a production table was permanently deleted by an agent with legitimate write access and no confirmation step for irreversible operations. Neither was a model-quality failure.

Wall 4 — Accountability: the question you cannot answer on a call

How many agents do we run, who owns them, what can they reach, and can you turn one off right now? The honest answer is usually a three-week-old spreadsheet. And this one now has a deadline: the EU AI Act's high-risk and transparency obligations began applying on 2 August 2026 — Article 12 (automatic event logging over the system's lifetime), Article 14 (human oversight), Article 19 (log retention), Article 26 (deployers must assign oversight to people with real competence and authority). None mandates an architecture. But “demonstrate effective human oversight of a high-risk decision” is a far easier claim to support with a system that can hold an action than one that can only show a log of what already happened.

Figure 1 — The same four agents, wired two ways

Point-to-point (where you are)

Agent A
Agent B
Agent C
Agent D
CRM
DB
Email
Payments

Policy lives in n×m places. Every new agent is a new integration, credential and audit gap. Turning one off means finding every place it is wired in.

Through a control plane

Agent A
Agent B
Agent C
Agent D
AGENT CONTROL PLANEidentity · policy · gateway · evidence · kill switch
CRM
DB
Email
Payments

Policy lives in one place and applies to agents built by teams you have never met. One revocation stops one agent everywhere, immediately.

Ungoverned pathMediated pathControl plane
Figure 1. The control plane is not extra architecture — it is the same architecture with policy, identity and audit factored out of n×m integrations into one mediated boundary. The identical argument that produced the service mesh, one layer up.

3. What it actually is

An AI agent control plane is a centralized management and governance layer that discovers, identifies, authorizes, observes and controls AI agents across their lifecycle, regardless of where those agents were built or where they execute.

Three clauses carry the weight. “Centralized” means policy is authored, versioned and evaluated in one place, not copy-pasted into each agent's system prompt — decisions centralize, enforcement distributes. “Across their lifecycle” extends scope past the request path: agents get registered, approved, versioned, suspended and retired. Most teams build for the request path only, then cannot answer “is this agent still supposed to exist?”

“Regardless of where they were built or execute” is the clause people skip, and it is what makes this infrastructure rather than a library. If your governance only works for agents on your framework, it will be routed around within a quarter — by a SaaS vendor's embedded agent, a partner integration, or one team that preferred a different SDK. A control plane that only governs compliant agents governs nothing.

The operational version of that definition is a test. A control plane is real to the extent it can answer these seven questions, on demand, from a system rather than from someone's memory.

Figure 2 — The seven-question test
1What agents exist?Including third-party, embedded and shadow agents nobody registered.
2Who owns each agent?A named human accountable for its behaviour, not a team alias.
3On whose behalf is it acting?The full delegation chain, preserved across every hop.
4What can it access?Data, tools, MCP servers, other agents — actual reach, not intended.
5What may it do?Permitted actions, evaluated per call, including irreversible ones.
6What did it do, why, at what cost?Trajectory, reasoning, policy decisions, spend — replayable.
7Can it be stopped now?Restrict, pause or terminate — including tasks already in flight.
1–2 = inventory3–5 = authority6 = evidence7 = control
Figure 2. Most organisations partially answer 1, 2 and 6. The gap concentrates in 3, 5 and 7 — delegation, per-action permission, and the ability to intervene. Those three are exactly what separates a control plane from a dashboard.

Warning

Demonstrate it, don't describe it. Pick one agent. Suspend it, revoke its credentials, terminate an in-flight task. Time it. If that needs a deploy, a config PR or a conversation with another team, you do not have a kill switch — you have an intention.

4. Control plane vs data plane

This split is why the architecture holds together, and it is the part most “governance layer” pitches leave implicit. The data plane is where work happens: agent runtimes, inference, retrieval, tool execution, agent-to-agent messages. The control plane is where decisions about that work are made: identity issuance, policy evaluation, budget enforcement, approval routing, evidence capture. They meet through two interfaces you will see in every serious design:

  • Policy Decision Point (PDP) — in the control plane. Takes a request context (who, what, on whose behalf, toward what goal, at what risk) and returns permit,deny, or permit with obligations like redaction or step-up approval.
  • Policy Enforcement Point (PEP) — in the data plane, inline on the action path. Asks the PDP and, critically, does not proceed until it has an answer.
Figure 3 — Decisions centralize, enforcement distributes
Control plane decides out of band
Registryinventory & owners
Identityissue & attest
PDPpolicy decisions
Budgetscost & quota
Evidenceaudit & replay
Data plane executes in the hot path
PEP: retrievalbefore context loads
PEP: tool / MCPbefore the call
PEP: delegationbefore the hop
PEP: outputbefore the reply
Model APIs
MCP servers
Systems of record
Other agents
Control plane — change without redeploying agentsFour enforcement boundaries
Figure 3. Four enforcement points, not one. An agent affects the world at exactly four boundaries — what it reads, what it calls, who it delegates to, what it emits — so a control plane covering only the tool call leaves three doors open.

Tip

Policy changes are the frequent event, not code changes. When legal says “no agent may send customer PII to a model hosted outside the EU,” you want that to be a policy edit that takes effect in seconds across every agent — including ones you did not write. If policy lives in prompts and if statements, the same request becomes a cross-team migration.

5. Reference architecture

Five horizontal layers: the consumers who initiate work, the agent runtimes that do it, the control plane that governs it, the mediated egress every action must pass through, and the resources that ultimately get touched. Read it top to bottom as the path of one request.

Figure 4 — AI agent control plane: reference architecture
1 · Consumers principals & triggers
Human usersSSO identity
Applicationsservice identity
Business eventsqueues, webhooks
Scheduled tasksno live human
2 · Agent runtime heterogeneous by design
First-party agentsyour frameworks
Vendor / SaaS agentsembedded, opaque
Multi-agent crewsplanner + workers
Shadow agentsdiscovered, not declared
3 · AI agent control plane the subject of this article
Agent registryinventory, owner, risk tier, version
Identity serviceper-agent workload identity
RBAC / ABAC engineroles, attributes, delegation chain
Policy engine (PDP)versioned, simulated, roll-backable
Guardrail servicesinjection, PII, unsafe sequences
Approval / HITL brokerstep-up for irreversible actions
Model routerroute by risk, residency, cost
Budget & quota governortokens, calls, retries, depth
Orchestration governortrust ceilings, loop limits
Context & memory governanceprovenance, masking, isolation
Observability & audit storetrajectories, decisions, evidence
Evaluation & release gatesadversarial tests, regressions
4 · Mediated egress the only way out
Tool & MCP gatewayallowlist, arg inspection, quarantine
LLM / model gatewayprovider auth, residency, spend
Agent-to-agent gatewaydelegation, scope attenuation

Network egress policy makes this path structurally unavoidable — agents cannot reach resources except through it.

5 · Resources where consequences are real
Systems of record
Databases & lakes
Vector stores
SaaS & partner APIs
Model providers
Control plane componentEnforcement point (fails closed)Discovered, not declared
Figure 4. Note what layer 2 admits: your agent estate will always be heterogeneous, and some of it opaque. That is the design constraint, not a mess to clean up. The control plane earns its place by governing agents it did not create — which is why enforcement sits at layer 4, where actions must pass regardless of how the agent was built.

6. “We already have IAM, a gateway and observability”

The most reasonable objection in this discussion. Most boxes in Figure 4 look familiar: identity is your IdP, RBAC is your entitlement system, the gateway is API management, observability is your APM vendor. So why is that insufficient? One sentence, and everything follows from it:traditional enterprise systems are deterministic, and AI agent systems are probabilistic.

Your platforms are not badly built. They are built on assumptions agents systematically violate. Read the right-hand column as a threat model, because that is effectively what it is.

Your platforms assume…But an AI agent…Consequence
Predefined execution pathsPicks tools at runtime and replans mid-taskYou cannot pre-declare the call graph, so you cannot pre-approve it
Dependencies declared in advanceDiscovers new tools and MCP servers dynamicallyDependency review happens after the dependency is in use
Permissions remain stable between releasesNeeds different authority per task, often per stepCoarse standing permissions become the permanent blast radius
Logic is inspectable in source codeHolds logic in weights, prompts, retrieved context and memoryCode review no longer tells you what the system will do
Failures are technical: latency, errorsFails semantically \u2014 a successful call toward a wrong goalEvery dashboard is green while the outcome is wrong
A request has one identified callerActs for a human, another agent or a process, several hops deepAttribution collapses into a shared service account
Services don't treat content as instructionsTreats documents, emails and tool responses as steering contextAny untrusted source becomes an instruction channel
Consumption is proportional to inputCan burn unbounded tokens and compute pursuing one objectiveA single task can generate an unbounded bill or runaway loop

Your API gateway can authenticate a caller and validate a payload — it cannot tell whetherthis call serves the goal the user asked for, or whether it is call four in a sequence nobody approved. Your IdP can issue a token — it has no concept of a chain running user → planner → worker with scope shrinking at each hop. Your APM can say a request took 900ms — not that the agent confidently chose the wrong tool and succeeded at the wrong thing.

7. The 10 building blocks

Each block below pairs what your existing platform does not do with what the control plane adds. You will not build all ten at once, and should not try — blocks 1–3 carry most of the risk reduction.

1Agent registry & lifecycle

Gap: CMDBs register services, not an agent's prompts, models, tools, knowledge sources, memory and delegated permissions.

Adds: discovery and registration of all agents including third-party and shadow ones; ownership, purpose, risk tier and dependencies; approval, deployment, suspension, versioning and retirement as first-class states.

2Identity & dynamic authorization

Gap: RBAC assigns predictable permissions to stable subjects. Agents act dynamically, for a user or another agent.

Adds: a distinct identity per agent; the full user → agent → agent delegation chain; short-lived task-scoped credentials; least privilege with step-up approval for sensitive actions.

3API, MCP & tool gateway

Gap: API gateways validate single requests without knowing the objective, prior actions, or the cumulative impact of many calls.

Adds: classified tool and MCP registry; per-agent allowlists; argumentand response inspection; unsafe-sequence detection; confirmation for irreversible operations; one-click quarantine of a compromised tool.

4Policy & guardrail management

Gap: policy engines evaluate structured attributes, not task intent, injected instructions or dynamically generated plans.

Adds: one place for model, tool, data, delegation, cost and geography policy; evaluation at all four boundaries; semantic risk detection paired with deterministic enforcement; versioning, simulation and rollback.

5Context, data & memory governance

Gap: data governance protects individual sources, missing the risk created when an agent aggregates, retains and forwards.

Adds: retrieval filtered by user and agent authorization; source provenance and trust levels; context minimization and masking; memory isolation across users and tasks; retention, residency and output DLP.

6Agent-aware observability & audit

Gap: monitoring measures latency, errors and availability. It does not explain decisions, tool selection or business outcome.

Adds: full task trajectories; users, agents, models, retrievals, tools and approvals correlated into one causal record; policy decisions and cost attached; outcome measurement; execution replay.

7Multi-agent orchestration

Gap: workflow engines follow predefined paths. Agents generate plans and delegate dynamically, so the graph does not exist until runtime.

Adds: governed discovery and delegation; trust ceilings so a delegate never exceeds its caller; depth, retry and time limits; loop and cascade detection; cancellation and compensation.

8Evaluation & quality

Gap: testing assumes determinism. Outcomes shift with model, prompt, context, memory and tools — five axes of drift your CI does not pin.

Adds: scenario-based and adversarial evaluations; tests for tool use, policy compliance and task completion; regression detection across versions; continuous production evaluation; release gates on risk thresholds.

9Cost & resource governance

Gap: FinOps reports cost after consumption. It cannot stop one agent task in real time — and that gap is where a runaway loop lives.

Adds: budgets per user, agent and task; limits on tokens, calls, retries and delegation depth; in-flight anomaly detection; model routing by task risk; circuit breakers.

10Runtime security & response

Gap: container and security platforms detect technical threats but cannot judge whether an action is appropriate for the assigned objective.

Adds: ephemeral sandboxes for risky tasks; restricted network, command and filesystem reach; injection, memory-poisoning and tool-abuse detection; agent suspension and credential revocation; preserved evidence.

Two of those blocks deserve a diagram, because they are the two most often implemented wrong.

Figure 5 — Effective permission is a conjunction (block 2)
User permissionwhat the principal may do
Agent permissionwhat this agent may ever do
Task permissionscope of this task only
Resource policywhat the target allows
Current risk policylive posture, incidents
ANDall five, per action
Permitexecute with scoped credential
Permit with obligationsredact, mask, require approval
Denyand deny by default on uncertainty
Any single false denies the actionAn unavailable input is not a true
Figure 5. The most common authorization bug in agent systems is treating this as an OR, or collapsing it to the first term alone. A user being able to issue refunds does not mean an agent acting for them may — and certainly not during a task about updating a shipping address.
Figure 6 — The unit of observability is the trajectory (block 6)
01Goal
02Retrievalpolicy check
03Decision
04Tool callpolicy check
05Delegationpolicy check
06ApprovalHITL gate
07Action
08Outcome
Control plane evaluates and can halt hereAgent-internal step
Figure 6. Four of eight steps are decision points where the control plane can stop the task. An observability tool records all eight and stops none — the entire distinction between evidence and control. Emit this with the OpenTelemetry GenAI conventions to keep it portable.

Tip

Registry (1) → Identity (2) → Gateway (3). Those answer five of the seven questions and give you a real enforcement point. Add cost governance (9) next because it funds itself, then observability (6) because it is what auditors and incident reviews actually consume.

8. The enforcement test

You will evaluate tools that all claim to be control planes. One question separates the real ones from very good logging, and it is not about features:

What happens when the policy check itself cannot complete?

If the action proceeds and the problem is logged, the system fails open. If the action is denied until the decision resolves, it fails closed. Fail-open is a debugging posture and a respectable one — it is not a control. Three properties distinguish enforcement from evidence-routing: it is a gate, not a queue (a review queue is fail-open for the interval before a human looks); it fails closed under uncertainty, because a denied action is recoverable while an unauthorized one may not be; and it is non-bypassable — atry/except around a policy call that defaults to allow looks like resilience and behaves like a backdoor.

Figure 7 — The same failure, two architectures
Agent requests an actionPOST /refund · amount 40,000
Policy check cannot completePDP timeout · unreachable · ambiguous
↓ fail-open path
Action executesthe refund is issued
Violation detected lateralert fires · review queue
Remediation is a reversalif reversal is even possible
↓ fail-closed path
Action is heldnothing has happened yet
Denied or escalatedagent gets an error it can handle
Remediation is a retrycheap, safe, reversible
Consequence already realConsequence still preventableUncertainty — the moment that decides everything
Figure 7. Both architectures produce an identical audit trail. Only one still has a choice. On the left your remediation options are reversal and apology; on the right, a retry.

This is a latency argument too. Benchmarking of closed-loop enforcement in multi-agent systems reported an observability-only baseline taking on the order of 15 seconds to surface a violation — long after the action completed — while inline enforcement held decisions under200 milliseconds and cut the share of violations that complete before anything stops them by a reported 97.6% (arXiv 2604.05119). A violation that routes regulated data to the wrong region does not wait for a dashboard refresh.

Warning

“Won't a fail-closed gate in the hot path cause outages?” It can, built badly. The mitigations are ordinary distributed-systems practice: cache policy and identity locally, decide in-process where possible, keep decisions to single-digit milliseconds, and scope fail-closed strictly toconsequential actions. Reads and idempotent calls can degrade gracefully; irreversible writes should not. Uniform strictness is itself a failure mode — Gartner has warned that applying identical governance regardless of an agent's autonomy and scope leads to failed deployments. Tier your agents and tier your enforcement.

9. What it is not

Most confusion in this category comes from adjacent products described with control-plane language. None of these is deficient — the table describes design intent, not quality. The value is knowing which question your stack actually answers before you tell an auditor it answers all of them.

CategoryDoes wellBlocks inline?Knows delegation?Full inventory?
Agent / LLM observabilityReconstructing what an agent did and whyNo \u2014 flags and queuesPartialNo
API gateway / API managementAuth, rate limits, schema validationYes \u2014 per requestNo \u2014 sees one callerNo
IdP / IAM & PAMHuman identity, SSO, entitlements, vaultingPartial \u2014 at auth timeNoPartial \u2014 service accounts
MCP / AI gatewaysTool federation, credential isolationYes \u2014 at the tool boundaryPartialPartial
Enterprise GRC platformsRisk registers, attestations, regulatory mappingNo \u2014 organizational layerNoPartial \u2014 self-reported
Agent frameworksBuilding agents: planning, tools, memoryNo \u2014 bypassable by designPartial \u2014 one frameworkNo
AI agent control planeDeciding whether an action is permitted, and proving what happenedYes \u2014 fails closedYes \u2014 core primitiveYes \u2014 incl. discovered

Two rows deserve a note. An MCP gateway is the closest single component to a control plane and often the right first purchase — but it governs one boundary and has no view of retrieval, delegation or output. An agent framework is the most tempting place to put policy and the worst place to rely on it: a check inside the agent's own code is removable by the agent's author, and it governs nothing built by another team.

Info

Observability is a prerequisite for a control plane, not a competitor. You cannot enforce a policy against a condition you cannot detect, nor produce evidence from telemetry you never captured. The mistake is treating the dashboard as the control.

10. Find yourself on this ladder

Figure 8 — Agent control plane maturity
L0Ad hocBorrowed credentials. Policy in prompts. Inventory is tribal knowledge. Nobody can answer question 7.
L1ObservedTracing in place; you can reconstruct what happened afterwards. Still fully fail-open: violations are alerts, not stops.
L2InventoriedRegistry with owners, purpose and risk tier. Per-agent identities replace shared accounts. Questions 1–3 answered.
L3EnforcedAll tool and model traffic mediated. Per-agent allowlists, fail-closed on consequential actions, step-up approval, working kill switch.
L4AdaptiveRisk-tiered autonomy: authority expands or contracts on live signals. Policy changes simulated before rollout. Evidence continuous and audit-grade.
Most enterprises sit at L1The hard jump is L2 → L3 — it needs a real enforcement point
Figure 8. L1 is a comfortable trap. It feels like governance because there are dashboards, and it passes most internal reviews. The transition that changes your risk profile is L2 to L3, where a decision starts being able to say no.

12. The core architectural principle

Tip

The AI agent control plane does not replace IAM, API management, observability, data governance or security platforms. It makes those systems agent-aware — by supplying the task context, delegation information, behavioural signals and continuous policy enforcement they were never designed to receive.

That framing turns the project from a rip-and-replace into an integration. You are not throwing out your identity provider; you are teaching it that an agent acting for a user is a different principal than either alone. You are not replacing your gateway; you are giving it the goal behind the request. You are not abandoning your observability investment; you are adding the one thing it structurally cannot do, which is say no.

And it reframes the opening question. “Do we need an AI agent control plane?” is not really a build decision. If you run more than one agent you already have one — scattered across prompts, environment variables, exception handlers and a spreadsheet, where nobody can inspect it, version it or trust it. The only decision in front of you is whether it becomes deliberate.


Frequently asked questions

What is an AI agent control plane?

A centralized management and governance layer that discovers, identifies, authorizes, observes and controls AI agents across their lifecycle, regardless of where those agents were built or where they execute.

Operationally it answers seven questions at runtime: what agents exist,who owns each one, on whose behalf it is acting,what it can access, what actions it may perform, what it did and at what cost, and whether it can be paused or terminated immediately.

What is the difference between a control plane and a data plane for AI agents?

The data plane is where work happens: agent runtime, model calls, tool invocations, data access. The control plane is where decisions about that work are made: identity issuance, policy evaluation, budget limits, approval routing, audit.

The control plane decides; the data plane moves. Separating them lets you change policy without redeploying agents, and enforce a decision even when the agent code was written by a team you do not control.

Why isn't IAM plus an API gateway enough?

Those systems were designed for deterministic software. They assume predefined execution paths, callers known at deploy time, dependencies declared in advance, and stable permissions.

An agent picks tools at runtime, discovers new ones, acts on behalf of a human or another agent, keeps memory across sessions, and can treat an untrusted document as an instruction. An API gateway can validate one request — it cannot tell that a sequence of individually legitimate calls adds up to an action nobody authorized.

Is it the same as an observability platform?

No. Observability tells you what happened; a control plane decides whether it is allowed to happen.

Most agent observability receives telemetry alongside or after execution, so a violated policy gets logged and escalated while the action continues. A control plane sits inline, holds the action until the decision resolves, and denies by default when it cannot decide. Observability is a required input, not a substitute.

Does it add latency?

Yes, but far less than the model calls it governs. A well-built decision is a local evaluation against cached identity and policy — single-digit to low tens of milliseconds — against inference measured in hundreds of milliseconds to seconds.

The right comparison is not zero latency versus some latency. It is a small inline cost versus discovering a violation after the action already completed.

What are the core components?

Ten building blocks: agent registry and lifecycle; identity and dynamic authorization; API, MCP and tool gateway; policy and guardrail management; context, data and memory governance; agent-aware observability and audit; multi-agent orchestration governance; evaluation and quality management; cost and resource governance; and runtime security with incident response.See all ten →

When do you need one?

The practical trigger is the second agent, the second team, or the first write action — whichever arrives first.

One read-only agent from one team needs good engineering, not a control plane. The moment two teams ship independently, or an agent can change a system of record, you need identity, permission and audit living outside the agent code.

References

Statistics and findings are summarised and paraphrased; follow the links for originals.

  1. Gartner, Six Steps to Manage AI Agent Sprawl (April 2026) — 150,000+ agents per average Fortune 500 by 2028. gartner.com
  2. Gartner, Uniform Governance Across AI Agents Will Lead to Failure (May 2026) — the 40% demotion projection. gartner.com
  3. A Five-Plane Reference Architecture for Runtime Governance of Production AI Agents, arXiv 2606.12320. arxiv.org
  4. AIP: Agent Identity Protocol for Verifiable Delegation Across MCP and A2A, arXiv 2603.24775 — MCP authentication scan. arxiv.org
  5. Governance-Aware Agent Telemetry for Closed-Loop Enforcement, arXiv 2604.05119 — detection-vs-enforcement latency. arxiv.org
  6. OWASP GenAI Security Project, Top 10 for Agentic Applications (2026). owasp.org
  7. Cloud Security Alliance, research on distinguishing agent from human actions (March 2026). cloudsecurityalliance.org
  8. Microsoft Security Response Center, CVE-2025-32711, and Aim Security's EchoLeak disclosure. msrc.microsoft.com
  9. SPIFFE workload identity, and OAuth 2.0 Token Exchange (RFC 8693) for delegation chains. spiffe.io · RFC 8693
  10. OpenTelemetry GenAI Semantic Conventions — portable agent-trajectory telemetry. opentelemetry.io
  11. EU Artificial Intelligence Act — Articles 12, 14, 19, 26; obligations applying from 2 August 2026. artificialintelligenceact.eu
  12. Databricks, reporting on multi-model-family adoption. databricks.com

Last reviewed 28 July 2026. Vendor categories, standards drafts and regulatory dates move quickly — verify dated claims against primary sources before relying on them in a design review.