Shadow AI Agents & Agent Sprawl: Identity Sprawl vs. Context Sprawl

Two agents, both provisioned through the same approval workflow, both carrying scoped credentials that would pass any identity audit clean, give two different answers to the same question: what was last quarter’s revenue. Neither autonomous agent did anything an identity review would flag. Neither violated a permission boundary. They just don’t agree on what’s true, and no agent governance program built around credentials and access control has any way to notice, let alone fix, that. This is what agentic AI adoption actually looks like at scale, not a hypothetical edge case reserved for the most immature deployments.

Agent sprawl gets treated, almost everywhere it’s written about, as a single inventory problem: too many agents, no registry, no clear owner. That’s real, and it’s half the picture, the same underlying inventory failure we’ve covered from the infrastructure side in MCP Gateway: The Solution to MCP Sprawl. The other half doesn’t show up in any credential audit, and it’s the half shadow AI agents actually expose once an organization has solved the easy part.

AI Agent Sprawl Is Actually Two Governance Problems, Not One

Gartner’s own numbers describe the scale of the inventory problem specifically. Speaking at the Gartner Digital Workplace Summit in London in April 2026, analyst Max Goss said the average global Fortune 500 enterprise will run more than 150,000 AI agents by 2028, up from fewer than 15 in 2025. Only 13% of organizations believe their AI agent sprawl governance is adequate today, a gap we’ve covered in more depth in our AI Agent Control Plane governance playbook, including how many organizations couldn’t shut down a rogue agent if one emerged. Kore.ai‘s own coverage of that research draws a useful distinction on top of it: agent sprawl is the governance and visibility problem, shadow AI is its downstream security consequence, and the relationship runs in one direction, unchecked sprawl produces shadow AI, not the reverse. Gartner’s own comment at the summit points at the same mechanism from a different angle: organizations that block agent use without a governed alternative just push employees toward ungoverned tools instead. In practice, agent deployment happens across different teams without a formal approval step, and IT teams are left reconstructing an inventory after the fact rather than governing it as it grows.

That distinction is correct as far as it goes. It’s also incomplete, because it treats agent sprawl as a single axis: how many agents exist, and does anyone govern them. Agent sprawl occurs on a second axis that inventory-and-permissions thinking doesn’t reach at all. Identity sprawl is uncontrolled growth in what an agent can authenticate as and what access it accumulated. Context sprawl is uncontrolled divergence in what different agents actually know, retrieve, and treat as ground truth. An agent can be flawlessly governed on one axis and completely ungoverned on the other, and most organizations currently building agent governance programs are only instrumented to see the first one.

Identity Sprawl, Briefly: The Overlap with Shadow IT

Identity sprawl is the dimension most existing coverage of agent sprawl already treats in depth, so this gets one section here, not a full rebuild. It’s the credential and permission problem: an agent’s own identity, what it authenticates as, what it accumulated access to over time, whether that access is still scoped to what the agent’s current task requires.

We’ve covered this mechanism in full elsewhere, including the specific finding that non-human identities, the broader category that AI agents and service accounts both fall under, outnumber human identities 144 to 1 in the average enterprise according to Entro Security’s H1 2025 research, up from 92 to 1 a year earlier. The overlap with classic shadow IT is real but partial: an unapproved SaaS subscription eventually surfaces through an expense report or a network scan; an agent’s accumulated credential sprawl usually doesn’t, because nothing about normal operation forces the same review. The full argument for why traditional identity governance built around human review cycles doesn’t transfer to machine identities, and what actually closes that gap, is in our guide to non-human identity sprawl. What matters for this article is narrower: identity sprawl answers “who is this agent, and what can it do.” It has nothing to say about the second question.

Context Sprawl: The Axis Identity Governance Can’t See, Even When Agents Access Sensitive Data

Atlan, a data catalog vendor, named this second failure mode directly in a piece published in mid-2026: two agents can both pass an identity governance check clean and still give completely different answers to the same business question, because nothing governs what data, documents, or semantic definitions either one is actually grounded in. That’s context sprawl, and Atlan’s own framing is precise about why it’s invisible to identity tooling specifically: it does not appear on any agent registry, because a registry answers questions about who an agent is, not what it currently believes. The stakes rise sharply once the agents in question access sensitive data rather than a revenue figure: two agents grounded in different versions of customer data, one from a record already corrected for a known error, one from the stale copy, don’t just disagree, they can each confidently hand a customer or a regulator a different answer about the same account.

The revenue example generalizes past Atlan’s own data-catalog framing. One agent’s “last quarter’s revenue” might be pulling from a stale nightly export. A second agent’s might be pulling from a live system that’s already been updated twice since the export ran. A third might be grounded in a spreadsheet someone manually corrected after the initial number was known to be wrong, with no signal to any other agent that the correction happened. All three agents can hold perfectly valid, correctly scoped credentials. None of them agree with each other, and an identity audit checking whether each agent’s access matches its assigned permissions will return a clean result for all three.

Why MCP Is Where Context Sprawl Becomes Concrete for AI Tools

Model context protocol is named for exactly this problem, standardizing how an agent retrieves context from external systems rather than requiring a custom integration per data source. Standardizing the transport does not standardize the content moving through it, and that gap is where context sprawl stops being an abstract data-governance concern and becomes a specific, addressable infrastructure problem.

Two MCP servers can both correctly implement the protocol and still hand connected agents genuinely different context: one wrapping a database replica that syncs hourly, another wrapping the live primary; one applying a row-level filter scoped to a specific business unit, another exposing the same table unfiltered because nobody configured the scope identically. Both servers pass every protocol-compliance check available. Neither is misconfigured in a way any access control audit would catch, because access control governs who can reach a server, not whether two servers claiming to describe the same underlying system actually agree with each other. This is the same underlying proliferation pattern covered in our piece on shadow MCP servers, just examined for data consistency rather than credential exposure. Context sprawl at the MCP layer is specifically this: a proliferation of servers each exposing a slightly different, ungoverned slice of “truth” about the same data sources, with no registry tracking which version of that truth any given agent is actually connected to at the moment it acts.

The Infrastructure Response

Closing identity sprawl and closing context sprawl require the same underlying discipline, centralization, applied to two different things, not a longer list of security policies nobody enforces at runtime. Proper governance at this layer means approved tools are the fast path, not a slower alternative to whatever a developer can stand up alone, the same argument we’ve made in full for the tool-level permissions specifically rather than server-level all-or-nothing access. Identity sprawl closes with centralized credential brokering and a real, queryable inventory of every agent identity, covered in depth in our guide to non-human identity sprawl. Context sprawl closes with the equivalent discipline applied to MCP servers specifically: a curated catalog where each server’s data source, scope, and owner are known, rather than a proliferation of ad hoc servers each quietly answering the same question differently. That’s what Obot, our platform, is built around: an Enterprise AI Control Plane with three jobs, define what agents can reach, choose where they run, and prove what they did. Each one maps to a part of both axes.

Gateway: Define What Agents Can Reach

Obot MCP Gateway is open-source (MIT license), self-hostable on Kubernetes or Docker, and also available as a managed service. Same product either way. It addresses both axes from the same control point rather than treating them as separate problems requiring separate tooling. Every MCP server an organization routes through it is visible in one catalog, which server exists, who owns it, what it’s scoped to reach, the identity half of the equation. The same catalog is where context sprawl gets decided rather than discovered: a team can approve one server per underlying system and document what it wraps, instead of letting three servers describe the same revenue table three ways. And because every tool call is logged at the request level, an investigator can see which server an agent was actually grounded in when it acted, the incident-response question a clean credential audit can’t answer. Governance as infrastructure-level guarantee, not per-server implementation bet: a gateway that only solved the identity question would leave the revenue-answer problem exactly as unsolved as it was before.

Device: Watch and Enforce With Obot Sentry

Both axes start with knowing which servers exist, and many don’t route through any gateway. MCP servers stood up locally in AI clients like Claude Code, Cursor, and Codex are exactly the ones nobody has full visibility into. Obot Sentry scans those client configurations on developer machines, reports which MCP servers, skills, and plugins each one loads, and can block servers that aren’t approved, so a fourth copy of the revenue data shows up in an inventory instead of in a conflicting answer.

Hosted: Claude Environments

Obot v0.27 added Claude Environments, hosted Claude Code environments that run under the same control plane. When agents run there, their MCP connections come from policy rather than each developer’s local config file, so every instance starts from the same approved set of context sources.

Where the Two Axes Compound Each Other

Neither Gartner’s single-axis governance framework nor Atlan’s data-catalog framing shows what happens when identity sprawl and context sprawl occur together, which is the more common case in practice than either occurring alone.

An identity-sprawled environment, ungoverned MCP servers standing up without a central inventory or clear ownership of who’s responsible for any given one, makes context sprawl actively worse. Nobody is positioned to notice that two servers are quietly serving inconsistent versions of the same data, because nobody has full visibility into whether the servers exist in the first place. Many organizations ask security teams to own this problem after an incident forces the question, when the actual fix belongs earlier, at the point a server gets stood up at all. Put plainly: you cannot govern the consistency of context you don’t know you have.

The reverse compounding matters just as much for incident response specifically. A context-sprawled environment makes identity incidents harder to investigate even when the identity side is fully governed. A clean credential audit tells an investigator what an agent was authorized to access. It says nothing about what that agent actually believed when it took a consequential action, which server’s version of the data it was grounded in at that moment, or whether that version was even correct. None of this is complicated individually. Together, across a fleet of MCP servers proliferating faster than any manual review can track, it’s a blind spot that compounds instead of one that simply adds up.

An organization can win the identity half of agent sprawl completely, a clean registry, scoped credentials down to the tool-call level, full audit logs, and still have agents confidently giving different answers to the same question. That’s not a failure of the identity program. It’s evidence the program was only ever built to answer one of the two questions agent sprawl actually asks.

FAQs

What is agent sprawl?


Agent sprawl is the uncontrolled, uncoordinated proliferation of autonomous AI agents across an organization, typically without a centralized registry, consistent ownership, or governance model. It commonly runs on two independent axes: identity sprawl, uncontrolled growth in agent credentials and access, and context sprawl, uncontrolled divergence in what different agents actually know and treat as true.

What’s the difference between agent sprawl and shadow AI?


Agent sprawl is the inventory and governance problem, how many agents exist, who owns them, what they can access. Shadow AI is a downstream security consequence: employees or teams routing around governed tools because sanctioned options feel too slow or restrictive. Kore.ai’s coverage of Gartner’s research frames the relationship as directional, unchecked agent sprawl produces shadow AI rather than the other way around.

What is context sprawl in AI agents?


Context sprawl is uncontrolled divergence in the data, documents, or semantic definitions different agents are actually grounded in, even when each agent’s identity and access are properly governed. Two agents can both pass an identity audit clean and still give inconsistent answers because nothing governs whether they’re working from the same version of the underlying truth.

How does Obot help with agent sprawl?


Obot MCP Gateway gives every MCP server a named owner, scope, and catalog entry and logs every tool call, Obot Sentry finds the servers configured locally on developer machines, and Claude Environments start agents from the same approved set of connections. Together they cover the identity axis directly and make the context axis something a team can decide and audit.

What is identity sprawl in AI agents?


Identity sprawl is the uncontrolled growth of an agent’s credentials, authentication methods, and accumulated permissions over time, the same underlying problem behind non-human identity sprawl generally, where machine identities outnumber human ones 144 to 1 in the average enterprise, according to Entro Security’s H1 2025 research.

How is context sprawl different from identity sprawl?


Identity sprawl answers who an agent is and what it’s authorized to access. Context sprawl answers what an agent actually knows and whether that knowledge is consistent with what other agents know. An identity audit cannot detect context sprawl, because context sprawl isn’t a permissions violation, it’s a data-consistency gap between systems an agent has every right to reach.

Can an agent be identity-governed but still context-sprawled?


Yes, and this is the common case, not an edge case. An agent can hold correctly scoped credentials, pass every access review, and still be grounded in a stale or inconsistent data source relative to other agents answering the same kind of question, because identity governance and context governance are enforced by different mechanisms entirely.

How does MCP relate to context sprawl specifically?


Model Context Protocol standardizes how agents retrieve context from external systems, but standardizing the transport doesn’t standardize the content. Two MCP servers can both correctly implement the protocol while wrapping different, inconsistent views of the same underlying data, a stale replica versus a live system, an unfiltered table versus a scoped one, producing context sprawl at the exact layer the protocol was built to standardize.

How do you reduce both identity and context sprawl at once?


Both require the same underlying discipline, centralized visibility, applied to different things: a real inventory of every agent’s identity and access closes identity sprawl, while a curated, centrally governed catalog of MCP servers with known, consistent data sources closes context sprawl. Treating them as one problem with one fix tends to solve only the identity half.