LetterMCP

Role of Platform Engineering in AI Agent Enablement

Platform engineers must rebuild identity, access.

Staff Writer · · 11 min read
Cover illustration for “Role of Platform Engineering in AI Agent Enablement”
Agentic AI Foundations · September 4, 2026 · 11 min read · 2,451 words

Platform engineering used to mean one thing: build the tools that let developers ship code faster. That job hasn't disappeared, but it's no longer the whole job. AI agents now sit on the platform as full-blown users. They call APIs, touch databases, read internal docs, and act on decisions nobody explicitly signed off on, and every one of those actions runs through the same auth, access, and credential systems a platform team already owns. Most companies built that scaffolding for humans, and treating it as good enough for agents is the mistake that's already causing breaches.

Why AI agents stress existing platform infrastructure in ways human users do not

Picture the newest hire on the platform. It doesn't sleep, it never skims the docs, and it can fire off a thousand API calls before a human teammate finishes their coffee. That's roughly the throughput gap between an agent and a person doing the same job through a UI.

Humans pause when something looks off. They message a coworker, they ask "wait, should I be doing this?" Agents follow instructions instead, and if the environment around them has a gap, they walk right through it without noticing there was ever a gap to begin with.

That difference breaks three assumptions built into most platforms, and treating any one of them as a minor gap is where teams get this wrong. Identity systems were built for people logging in with SSO; software needs its own scoped, short-lived credential it can rotate or kill on demand. Access control was built for humans you can hold accountable when something goes sideways; hand an agent that same broad access without narrowing it down, and every tool it touches sits exposed at full privilege. Observability was built around the trails people leave without trying, tickets, chat threads, approval requests. Agents leave none of that unless someone builds the plumbing to capture it on purpose.

Model Context Protocol, or MCP, has become the standard way agents plug into enterprise systems, and its growth explains why this matters now instead of later. The public MCP server registry grew from roughly 1,200 entries in early 2025 to over 9,400 by mid-April 2026, a sevenfold jump in attack surface in about fourteen months. Real-world agent incidents traced in security research point repeatedly to identity, tool access, memory, and supply-chain weaknesses, not to the underlying model. That's the exact territory platform teams already govern. Stretching a human-built identity system to cover agents requires a rebuild, and most teams haven't started one.

Diagram: MCP Attack Surface: A Sevenfold Jump in 14 Months. Visualizes: Show the growth of the public MCP server registry from roughly 1,200 entries in early 2025 to over 9,400 by mid-April 2026 -- a sevenfold increase in about fourteen months.

The security exposure that emerges when no platform foundation exists

Security researchers have counted close to 7,000 internet-exposed MCP servers, and roughly half had no authentication at all. That's what happens when teams stand up agents on their own, with no shared foundation underneath them, because building one takes longer than shipping the demo. Skipping the foundation is the decision that produces the incident.

Three failure modes show exactly what's at stake when that foundation is missing.

Prompt injection, cataloged as ASI01 in the OWASP framework, lets an attacker hide instructions inside a document or webpage an agent reads. The agent treats those hidden instructions as legitimate and acts on them, no malware required. EchoLeak, tracked as CVE-2025-32711, showed this in the wild: a prompt buried inside an email triggered Microsoft 365 Copilot to leak data the moment it processed that email automatically. The lesson runs deeper than one bug. The trust boundary sits inside every piece of text an agent reads, not inside the agent's own sandbox. On the supply-chain side, the mcp-remote proxy carried a command-injection flaw, CVE-2025-6514, that touched more than 400,000 downloads and turned any unpatched install into a door for stealing API keys and SSH credentials.

Overprivileged access compounds all of it. Configure one MCP server with database write access, and every agent connected to it inherits that power, including the ones that only needed to read a single table. Least-privilege design collapses fast when servers get built for maximum flexibility instead of tight scope, and that's the default most teams reach for, because it's easier to ship.

Security researchers consistently find that many organizations lack the visibility to track what data their agents touch, and agents routinely act outside their intended scope before anyone notices. That gap between what agents are doing and what anyone can see them doing is exactly where incidents start piling up. Every one of these failures traces back to a missing platform control: no scoped credentials, no policy enforcement, no audit trail. Fix the control, and the failure mode goes away; there's no version of this that gets solved by a better model.

What platform teams actually need to build: the shared infrastructure for safe agent deployment

Diagram: Four Controls Every Agent Platform Needs. Visualizes: Visualize the four platform controls platform teams must own: (1) Non-human identity -- per-agent, just-in-time credentials that expire when the task closes; (2) Tool gateway --…

State the design principle plainly: agents should travel the same approved route as human developers, through the same compliant modules, the same cost guardrails, the same security defaults, just automated instead of manual. Anything short of that means every team reinvents the same wheel, badly, under deadline pressure.

Four pieces make up that route, and platform teams need to own all four outright. Delegating pieces to whichever team ships an agent first is how you end up with five incompatible versions of the same control by next quarter.

Non-human identity comes first. Every agent needs its own identity, not a shared service account passed around five different automations. Credentials should get issued just-in-time and expire the moment the task window closes; that's what kills off the long-lived tokens that are so easy to steal and reuse. Agent identities should flow through the same SSO and SCIM lifecycle as human accounts, provisioned and deprovisioned the same way.

A tool gateway comes next. A centralized gateway should sit between every agent and every tool call, enforcing policy at the level of the individual call, not just at the network edge. Agents should only get access to the MCP servers their specific task requires, never the full catalog by default.

Credential management belongs in one place, not scattered across a dozen teams storing API keys in environment variables or config files. Rotation and revocation need to work per-agent and per-task, not just at the account level, and no agent should ever share a credential with another. Every action has to trace back to one distinct identity, or an audit becomes guesswork.

Governance rails finish the set. Policy gets defined once, centrally, and applies automatically the moment any team deploys an agent through the platform's approved path. Compliance rules around data residency should get built into that path directly, so a team can't accidentally break a regulation nobody told them about. Someone also needs one dashboard showing every MCP server, every skill, every agent live in the environment at once; you can't govern what you can't see.

Teams that build these four pieces once stop rebuilding the same plumbing every time another team wants to ship an agent. That's the same payoff golden paths already delivered for developers: when the safe path is also the fast path, people take it without being told to.

Observability and audit as load-bearing parts of the platform, not reporting afterthoughts

Many companies can't track what their agents touch. That's not just a security gap. It means there's no way to investigate after something goes wrong, and no way to show an auditor what actually happened, which turns every incident into a guessing game.

Human developers leave a trail without trying: approvals, tickets, chat logs. Agents leave a trail only if someone built the system to capture one, and skipping that step is a blind spot baked in from day one. It's the single most common gap platform teams find once they go looking, and it's usually the first thing that surfaces the moment something breaks.

Real observability for agents means full tracing across every tool call an agent makes, start to finish, not a summary showing only the beginning and the end. It means audit logs that can't get altered after the fact and that tie every action back to a specific identity, detailed enough to hold up in a forensic review or in front of a regulator. And it means real-time detection tuned to MCP-specific attack patterns, tool poisoning, prompt injection, intent drift, data exfiltration, caught at the level of the individual tool call rather than the session.

Intent drift deserves its own callout, because it's sneaky by nature. An agent can start a task well within its approved scope and slowly wander into territory nobody signed off on, one reasonable-looking step at a time. Catching that means comparing what the agent is actually doing against what it was supposed to do, and that comparison only happens if the platform runs it automatically. The OWASP Agentic Top 10's categories for multi-agent failure modes depend on catching anomalous behavior before it spreads through a multi-agent workflow. Observability is the only thing that makes that catch possible. Get this right, and security stops reacting to incidents after the fact and starts catching them mid-flight.

How the golden path model extends from developer enablement to the rest of the enterprise

The golden path idea from platform engineering, one approved route that's genuinely faster than any workaround, doesn't stop at developers. Non-technical employees want to automate their own work with agents too, and most of them have zero safe infrastructure to do it through. That's the gap platform teams keep missing, because they're still framing this as a developer-tooling problem when it's already a company-wide one.

When there's no golden path for them, people go find their own tools. They plug agents into enterprise data through whatever consumer integration is closest at hand, and that's precisely how shadow MCP usage starts: the kind security teams can't see coming and can't shut down after the fact.

The fix isn't complicated in concept. Take the platform team's existing infrastructure, centralized auth, scoped credentials, policy enforcement, and put a simple front end on it: a Slack-based setup flow, a catalog of tools that are already pre-approved. Nobody outside the platform team needs to see the plumbing underneath. What makes that catalog actually usable, though, is size. A narrow list of five approved MCP servers won't cut it if someone's task needs a sixth, and the moment it doesn't cut it, they'll go around it. The breadth of what's pre-vetted decides whether people stick to the governed path or wander off it.

Adoption follows the path of least resistance, full stop. Make the safe path the easy one, and usage concentrates there instead of scattering into a dozen ungoverned corners. Do this well enough, and security stops carrying the reputation of the department that says no. It becomes the reason the rest of the company can say yes.

Where platform engineering sits in the enterprise AI governance stack

Most organizations agree governing AI agents matters; far fewer have any actual policy in place. Companies aren't confused about whether this matters. What's missing is the infrastructure that would let a policy get enforced instead of just written down and filed away, and that gap is an infrastructure failure, not an awareness one.

The OWASP Agentic Top 10 sits across three layers at once: the model layer, covered by the separate LLM Top 10; the behavior layer, the Agentic Top 10 itself; and the connection layer, where MCP security concerns concentrate. Platform engineering owns the connection layer and most of the behavior layer, and that's where the majority of documented incidents actually start, not at the model. Put the money and attention on model safety alone, and you've secured the layer that isn't where the breaches happen.

What platform engineering brings to governance that no other function can replicate: enforcement built into the infrastructure itself, sparing teams from relying on a policy PDF someone's supposed to read and remember. Audit trails that are attributable by design, so compliance and legal get what they need without asking every product team to bolt on its own logging. One inventory of every agent, every MCP server, every tool live in the company, which is the baseline requirement for any real risk assessment.

Skip building this centrally, and the alternative carries real cost. Every product team ends up building its own version of auth, access, and audit. Governance turns inconsistent from one team to the next, and the company winds up with as many different security postures as it has teams deploying agents. Get this layer built right instead, and platform engineering becomes the thing that makes AI governance actually enforceable across the whole company, not just on paper. Security gets visibility, compliance gets an audit trail it can trust, and business teams get a path to shipping an agent that doesn't require a bespoke security review every single time.

Platforms designed for 2026's agentic workloads need to cover the full OWASP Agentic Top 10 surface, extending well past the access control problems identity platforms were originally built to solve. Money is already moving into this layer. The real question for any given company is whether its own investment is actually scoped for agentic workloads, or whether it's still funding tools built for a world before agents showed up.

What a platform team needs to assess and prioritise first

Start with inventory, because this isn't a future problem to plan for, it's a present one to find. Which agents are already running somewhere in the company? Which MCP servers do they talk to? Which of those connections actually have authentication behind them? Ungoverned agents are very likely already live in production, not something scheduled for next quarter's roadmap.

From there, three areas deserve attention first, in order of where the documented risk actually sits.

Authentication gaps come first. Check which MCP servers in active use have no auth controls at all, remembering that roughly half of internet-exposed servers carried none as of early 2026, and flag any long-lived shared credentials that can't be scoped or pulled back per agent.

Privilege scope comes next. Find every agent running with broader access than its actual task requires, since that overprovisioning is exactly what turns a small mistake into a serious breach.

Observability comes third. Figure out which agent actions leave no trace right now, because that's the blind spot an incident will eventually find on its own, whether the platform team is ready for it or not.

None of this calls for guesswork. The failure modes are already documented, the fixes are already understood, and the infrastructure to prevent them is largely the same infrastructure platform teams have spent years building for developers. It just needs to point at a different user now, one that doesn't sleep and doesn't ask permission before it acts.

Sources

  1. labs.cloudsecurityalliance.org

More in Agentic AI Foundations