Google announced general availability of the Gemini Enterprise Agent Platform this week, and buried in the release notes is the feature that actually matters to anyone running AI agents in production: every agent now gets its own cryptographically-attested identity, built on the open SPIFFE standard, bound directly to the agent's runtime, with every action it takes logged in a way nobody can quietly edit after the fact. Google is calling it Agent Identity. What it really is, is an admission that the industry's default way of running AI agents — hand them a shared API key, give them broad access because scoping things properly takes an extra afternoon, and hope nothing goes wrong — was never going to survive contact with agents operating at real business volume.
For a business that has spent 2026 watching one AI agent security incident after another make headlines — sandbox escapes, credential sprawl, agents quietly reaching systems they were never supposed to touch — a major cloud vendor shipping native, mandatory identity infrastructure for agents sounds like exactly the fix that's been missing. It mostly is. But the fix only covers agents running on Google's platform, and for most businesses right now, that is not where most of their agents actually live.
What Agent Identity actually does
Each Gemini Enterprise agent gets a unique SPIFFE ID mapped to its resource URI, strongly attested and tied to its lifecycle — so a stolen token can't be replayed by a different agent pretending to be the original. Access is enforced least-privilege by default rather than granted broadly and trimmed later. Every operation is logged non-repudiably, and dormant credentials get cleaned up automatically instead of sitting unused and unaudited for months. A new Agent Gateway sits in front of the whole ecosystem, enforcing access rules and screening for prompt injection and tool poisoning before an agent ever touches a governed resource.
Why This Is the Right Fix, Built at the Right Layer
The reason agent identity has been such a persistent gap isn't that businesses don't care about security — it's that identity and access management was built for people. A human employee gets onboarded once in an HR system, gets a login, and gets deprovisioned once when they leave. An AI agent doesn't fit that model at all. It can be spun up by a script, chain a dozen tool calls together in one task, and — because it reasons about which tool to call next rather than following a fixed integration path — its effective behavior can shift without anyone touching a config file. Bolting traditional IAM onto something that dynamic never worked well, which is exactly why so many teams skipped it and used a shared, over-scoped service account instead.
Building identity as a first-class, non-human-native concept — a SPIFFE ID tied to the runtime itself, not a password an agent happens to hold — is the structurally correct answer to that mismatch. It's also a strong signal about where the rest of the industry is heading: when a platform with Google's enterprise footprint makes this the default rather than an optional add-on, it stops being a nice-to-have security posture and starts being the baseline businesses will be expected to meet.
The Catch: It Only Governs What Runs On It
A realistic scenario
A mid-sized logistics company we advise runs its customer-facing support agent on Gemini Enterprise — so that agent now inherits Agent Identity's scoped credentials and audit logging automatically, at no extra engineering cost. But the same company also runs a dispatch-automation agent built directly against their TMS API using a custom Python service, and a document-processing agent built on a separate no-code platform for invoice extraction. Neither of those agents touches Google's identity layer at all. They still authenticate the way they always did — with long-lived API keys checked into a deployment config, no per-agent audit trail, no automatic credential expiry. The company's most-governed agent and its least-governed agent now sit three security tiers apart, and nobody planned it that way; it's just where each agent happened to get built.
That's the pattern most businesses are actually going to hit. Very few companies run every AI agent on a single vendor's platform, and the ones that do usually got there by accident rather than design — a support bot here, a workflow automation there, a document pipeline built by whichever team needed it fastest. Agent Identity is a real, well-built solution to a real problem, but it solves that problem for a fraction of the agents most businesses actually have running. The governance gap doesn't close. It just moves to whichever agent wasn't built on Google's stack — which, for a lot of companies, is most of them.
What This Means for Your Agent Rollout
This is a good moment to check where your own agents actually stand, because the answer for most companies is somewhere between "we haven't audited this" and "we know it's messy but haven't had time." A few things worth doing regardless of which platforms your agents run on:
- 1Inventory every AI agent in production and note what credential each one runs on — a scoped, agent-specific identity, or a shared key borrowed from an existing integration
- 2If you're already on Gemini Enterprise, confirm Agent Identity is actually enforcing least-privilege scopes on your agents rather than running with legacy broad permissions carried over from setup
- 3For agents built outside a governed platform — custom API integrations, no-code tools, agents on other cloud providers — apply the same principles manually: unique credentials per agent, scoped access, logged actions, automatic expiry on unused tokens
- 4Don't let "our main platform now has good governance" become the reason the other 60% of your agents never get audited
Where Wizeb comes in
Every custom AI agent Wizeb builds ships with scoped, agent-specific credentials and a real audit trail from day one — whether it's running on Gemini Enterprise, a different cloud stack, or a fully custom integration against your existing systems. We don't assume the platform's default security posture is enough, and we don't wait for a vendor announcement to tell us identity matters. If you're not sure whether your current agents would pass the audit described above, that's a conversation worth having before an incident forces it — start at wizeb.com/services/ai-agents.
Google building real identity infrastructure for agents is a genuinely good development, and it's a strong signal the rest of the industry will have to follow. But "our platform vendor solved this" is only true for the agents actually on that platform. For everyone else — which right now is most businesses running more than one kind of AI agent — the governance work is still yours to do.