The headline numbers from OutSystems’ April 2026 State of AI Development survey are being cited everywhere, usually in celebration: 96% of enterprises are already running AI agents; 97% are exploring organisation-wide agentic strategies.
What the celebrations omit is the number buried three pages deeper in the same report. Only 12% of those organisations have implemented centralised platform management for the agents they are running. Ninety-four percent report that agent sprawl is already “increasing complexity, technical debt, and security risk.”
This is not an adoption story. It is an agent sprawl story — the agents are live, the governance is not.
What Agent Sprawl Actually Looks Like at 500 People
Agent sprawl is not a metaphor. It has a specific technical signature that becomes legible when you audit a mid-size engineering organisation that has been shipping AI features for twelve to eighteen months.
A typical 500-person tech company that began deploying agents in 2025 will have, by mid-2026: a customer-facing support agent provisioned by the CX team; two to five internal workflow agents built by separate squads on different orchestration frameworks (often LangChain, CrewAI, or AutoGen — rarely the same one); a code-review agent in the CI pipeline; one or more data-pipeline agents on a schedule; and an indeterminate number of “experimental” agents piloted by individual contributors and never formally decommissioned.
Gartner’s April 2026 analysis, “Six Steps to Manage AI Agent Sprawl,” frames the trajectory starkly: the average Fortune 500 enterprise will have over 150,000 agents in use by 2028, up from fewer than 15 in 2025. Aligning enterprise controls with the NIST AI Risk Management Framework and CISA Guidelines for Secure AI System Development is becoming critical as governance lags behind adoption.
Four specific failure points define what agent sprawl looks like operationally.
Identity fragmentation. Each agent at deployment requires credentials: API keys, OAuth scopes, service-account roles, cloud IAM permissions. In practice, these are provisioned under sprint-cycle pressure, over-scoped for convenience, and rarely revoked.
Strata’s 2026 analysis of enterprise agent deployments found that agentic tokens are actively exposed across Jira tickets, Teams messages, Confluence pages, and code commits — not as a hypothetical future risk, but as present-tense reality. The model for human IAM, where roles drift slowly and audits can catch up, does not translate. Agent privilege escalation, as Strata documents, “accelerates” and occurs at development velocity. A service account provisioned for a now-deprecated pilot agent does not disappear from the identity plane when the agent does.
The technical fix is available but not widely deployed: ephemeral, task-scoped credentials that expire when a workflow completes, cryptographic identity verification via OAuth Dynamic Client Registration (DCR), PKCE flows, or SPIFFE/SVID certificates, and runtime policy enforcement through identity-aware proxies. According to Strata’s 2026 research, fewer than one in five organisations has implemented all three layers consistently.
Audit trail gaps. When an AI agent takes an action — sends an email, modifies a database record, calls an external API — what log does the organisation have? In most current deployments: an application log of the action, with no record of the reasoning chain that produced it, the tool calls that preceded it, or the identity of any parent agent that delegated to it. This is adequate for debugging. It is not adequate for compliance, incident response, or liability attribution.
The compliance requirement is becoming explicit. Microsoft’s Cloud Adoption Framework for AI agents, updated in early 2026, specifies that audit logs must record both identities — the agent and the user it is acting on behalf of — so that an auditor querying by user can reconstruct everything done in their name across every agent workflow. Sub-agents must authenticate as themselves, not inherit a parent’s credentials. Very few current agent deployments satisfy this requirement.
Cost attribution collapse. Enterprise AI token consumption has increased approximately 13x since January 2025, according to spend analysis from Elvex. That growth is not proportional to user growth. It is driven by agentic workflows — multi-step reasoning loops, tool-call retry sequences, and long-context operations — that compound token consumption in ways that quarterly forecasts built around per-seat pricing models cannot predict.
$6M+
Unplanned AI spend run up by one healthcare enterprise in six months — before finance traced the cause, per Elvex.
The root causes were runaway retry loops, long-context accumulation across multi-turn workflows, and the absence of per-agent budget controls — costs invisible in existing dashboards because spend was aggregated at team level, not attributed per agent.
Failure mode misclassification. Production agent failures sort into recognisable categories: context blindness, rogue actions, silent degradation, memory corruption, and runaway execution. Analysis of enterprise deployments consistently finds that the large majority trace to infrastructure gaps — missing guardrails, permission misconfigurations, absent monitoring — rather than model capability. Most organisations direct post-failure engineering at the model layer. They are debugging the wrong thing.
The Observability Tooling Gap
The observation that “89% of organisations running agents in production have implemented some form of observability” (LangChain, State of Agent Engineering 2026) should not be reassuring. “Some form” does a great deal of work in that sentence.
The critical distinction is between monitoring and observability. Monitoring tells you an agent failed. Observability tells you which step in the reasoning loop caused the failure, what the intermediate state was, and whether it was retryable. A general-purpose APM dashboard provides the former; practically none, without significant custom instrumentation, provide the latter.
The tools built specifically for this problem divide into two categories by heritage. ML-heritage platforms — Arize Phoenix, Weights & Biases Weave — bring evaluation rigour from the model development world and extend it into production tracing. Arize’s enterprise platform, Arize AX, is built on the Phoenix open-source foundation and treats traces as first-class AI artefacts with tool-call fidelity tracking, multi-step reasoning attribution, and LLM-as-judge evaluation scoring on production samples. W&B Weave is best suited for teams already using the Weights & Biases experiment-tracking ecosystem; its agent-specific debugging features are newer and still maturing relative to core model monitoring.
Infrastructure-heritage platforms — Datadog LLM Observability, Honeycomb LLM Observability — integrate cleanly with existing engineering stacks but carry the design assumptions of their origins. Datadog’s LLM Observability module surfaces token counts, error rates, and latency; it has no native understanding of prompt chain semantics or tool-call fidelity. Honeycomb’s event-based architecture makes it well-suited to deep tracing of individual agent workflows but requires more custom instrumentation to produce the cross-agent cost attribution and failure classification that governance-focused use cases require.
The standardisation layer that would allow organisations to instrument once and route to any backend exists in principle. OpenTelemetry’s GenAI Semantic Conventions SIG is defining attribute schemas for LLM calls, agent invocations, and tool executions, and major vendors including Datadog, Honeycomb, and New Relic already claim support.
The catch: as of early 2026, every category within the GenAI semantic conventions is still marked Experimental. There is no stable, agreed-upon standard for what agent observability should capture. Organisations instrumenting production agents today are making bets on schemas that may change.
The governance-relevant gap is specific: none of the current platforms natively connects observability traces to agent identity records, cost attribution ledgers, and failure-classification taxonomies in a single queryable view. These remain three separate data problems, addressed by separate tooling, requiring manual correlation for any governance use case that spans all three.
Three Agent Sprawl Governance Patterns That Work
The organisations that have navigated this are not waiting for a consolidated platform. They are implementing architectural patterns that work with current tooling constraints.
Pattern 1: The Agent Registry as a System of Record
The prerequisite for all other governance is a centralised, continuously updated inventory of every agent in the environment. Gartner’s April 2026 governance framework identifies this as step two of six — second only to establishing governance policies — precisely because you cannot govern what you cannot enumerate.
The technically harder version of this is discovery, not just registration. Arthur AI’s 2026 analysis of enterprise agent governance landscapes draws a sharp distinction: a registry catalogs agents that teams report; a discovery platform identifies agents that teams did not report. The highest-risk agents in most enterprise environments are the ones in neither inventory.
Continuous automated detection via OpenTelemetry scanning, MCP gateway analysis, and network traffic inspection is the only mechanism that catches shadow agents — agents deployed by individual contributors, third-party vendors, or automated pipelines outside the sanctioned build process.
EY’s documented buildout of an enterprise-scale agentic AI operating system for its 400,000-person global workforce illustrates the required scope: a centralised model catalog, policy enforcement at the gateway layer, responsible AI guardrails applied uniformly, and multi-agent workflow governance under a single management plane. The architecture is replicable; the key design decision is whether the registry is a passive catalog or an active enforcement point.
Pattern 2: Identity as Infrastructure, Not Configuration
The shift required is treating agent identity with the same operational rigour as production infrastructure: provisioning through infrastructure-as-code, testing in pre-production identity simulation environments, and automatic expiry when the agent lifecycle ends.
Aembit’s IAM-for-agentic-AI framework, updated in 2026, specifies the minimum viable architecture: per-agent unique registered identities (not shared service accounts), scoped OAuth grants that expire at workflow completion (not persistent credentials), delegation via scoped permission tokens when agent A spawns agent B (not credential propagation), and runtime policy enforcement through an identity-aware proxy that can revoke access without redeploying the agent. The proxy layer is the critical element — it makes identity enforcement a live control rather than a deployment-time configuration.
The 2026 Data Breach Investigations Report (Verizon) confirmed the vector: credential compromise affecting non-human identities — service accounts, API keys, OAuth grants — represents a growing and increasingly prominent initial access method in enterprise breaches, with Verizon explicitly flagging machine and service accounts as “the ones likely to be leveraged in our potential agentic AI future.” The remediation architecture is not novel; it is standard zero-trust applied consistently to a new identity type that most organisations have been treating as second-class.
Pattern 3: Per-Agent Budget Enforcement with Circuit Breakers
Cost attribution at the agent level is achievable with current tooling, but it requires connecting three data sources that are typically managed separately: the agent’s registered identity record, the token-level billing API from the model provider, and the observability traces that explain cost spikes behaviourally.
The implementation pattern documented by Prefactor (a cost-attribution platform for AI workloads) and corroborated by TrueFoundry’s AI gateway architecture (which enforces per-agent budget limits at the request layer) involves tagging every API call, tool invocation, and token consumption event with the requesting agent’s registered identity. Downstream, per-agent dashboards show cost broken out by agent, team, project, and business unit. Per-agent budget limits, configured as hard thresholds rather than alerts, automatically throttle or suspend agents that exceed allocation.
The circuit-breaker component addresses the runaway-loop failure mode directly. When token consumption per workflow transaction exceeds a defined threshold, the circuit breaker freezes the agent’s state and escalates to a human-in-the-loop queue. Not all failures are retryable — the classification matters. A tool-call failure caused by an unavailable API is retryable after backoff. A goal-drift failure requires human review before any retry. Organisations applying uniform retry logic to all failure types are compounding both cost and risk.
The 12-Month Outlook for Agent Sprawl
Several structural changes make mid-2027 materially different from mid-2026.
The Model Context Protocol (MCP), now approaching broad adoption as the standard interface layer between agents and tools, is becoming the natural insertion point for governance controls. MCP gateways that enforce per-tool access policies, log every agent-to-tool interaction with identity context, and apply rate limits and budget constraints at the protocol layer — rather than requiring application-layer instrumentation — are already in production at a small number of organisations and will become standard infrastructure within twelve months.
The OpenTelemetry GenAI semantic conventions will exit Experimental status during this period. When they do, the current fragmentation between observability backends resolves into a portable instrumentation layer — instrument once, route to any compliant backend. This substantially reduces the per-deployment tooling cost of agent observability.
The identity problem will be driven toward resolution by regulatory pressure rather than architectural preference. The EU AI Act’s high-risk system provisions, now in full enforcement, require documented accountability chains for automated decision systems. An agent that modifies customer data or executes financial transactions without a verifiable, attributable audit trail does not satisfy those requirements, compounding the legal risks detailed in our review of AI agent liability frameworks. Legal exposure, not best practice, will compel the IAM architecture upgrades that governance advocates have been recommending since 2024.
- 96% of enterprises run AI agents; only 12% have centralised platform management for them.
- Identity fragmentation, audit trail gaps, cost attribution collapse, and failure mode misclassification are the four operational signatures of agent sprawl.
- Three working patterns exist today: an agent registry with active discovery, identity treated as infrastructure with runtime policy enforcement, and per-agent budget enforcement with circuit breakers.
- Regulatory pressure (EU AI Act) and MCP’s emergence as a governance insertion point will force the architecture within 12-18 months, not best practice alone.
The Gartner finding that only 13% of organisations believe they have adequate agent governance is a more useful benchmark than adoption rates. By mid-2027, that number will move — not because agent sprawl disappears, but because the cost of not addressing it will be quantifiable in breach reports, budget overruns, and compliance findings.
The 96% adoption figure will keep rising. The governance infrastructure will follow — on a lag of twelve to eighteen months, and under duress. The organisations that build it now are not ahead of a trend. They are ahead of a crisis.
This article was produced with AI assistance and reviewed by the editorial team.
—