An agent is about to open a customer record. It has a plan, a tool, and a permission check. The only open question left is whose name goes on the request.
For most teams shipping agents this year, the answer has been a shared service account with broad scopes. Google changed that answer on November 13, 2026, when three-legged OAuth (3LO) with the Agent Identity auth manager reached general availability for Agent2Agent (A2A) agents in Gemini Enterprise.
Primary source: Gemini Enterprise release notes (opens in a new tab).
What shipped
Three moving parts landed together.
| Part | What it does | Where it runs |
|---|---|---|
| 3-legged OAuth (3LO) | The agent authenticates to a downstream system as the signed-in user | A2A agents built with the Agent Development Kit (ADK) on Agent Runtime |
| Agent Identity auth manager | Validates the consent callback and finalises the delegated credential in a Google-managed vault | Gemini Enterprise |
| Agent Registry bindings | Resolve which auth provider serves a given agent, MCP server, or API | Configuration layer |
The flow, in order:
- An A2A agent needs a downstream MCP server, a Google Cloud service, or a third-party API.
- If that resource requires consent, Gemini Enterprise prompts the signed-in user.
- The callback is validated, the credential is finalised by the Agent Identity auth manager, and the A2A conversation resumes on its own.
- The SDK carries the pieces:
pip install --upgrade "google-adk[a2a,agent-identity,mcp]"withgoogle-adk>=1.29.0.
The identity model underneath pairs a cryptographically attested SPIFFE agent identity with scoped, user-delegated 3LO credentials. Reference docs: Agent Identity (opens in a new tab) and 3-legged OAuth for A2A agents (opens in a new tab).
Why it matters: the credential is the control plane
Once an agent borrows the user's identity, three things become true at once. Scope shrinks to what that person can already reach. Every tool call carries a name. Revoking access is a single action on one account.
Google shipped the same idea at the browser layer. Chrome Enterprise now exposes GeminiActOnWebSettings (0 allowed, 1 blocked) alongside GeminiActOnWebAllowedForURLs and GeminiActOnWebBlockedForURLs, which scope exactly where an in-browser agent may act. When a URL matches both lists, the allowlist wins. GenAiDefaultSettings sets the organisation-wide default, and GeminiEnterpriseSettings binds a managed profile to a Gemini Enterprise project. Policy reference: Chrome Enterprise policy list (opens in a new tab).
Read the two releases together and the pattern is the same one, twice: an agent inherits a narrow, named, revocable grant, and the boundary is written down somewhere an admin can read.
user signs in
|
v
+---------------------------+
| signed-in user (3LO) |
+-------------+-------------+
| delegated, scoped
v
+---------------------------+
| A2A agent on Agent | <- attested agent identity (SPIFFE)
| Runtime (ADK) |
+-------------+-------------+
| tool call + audit line
v
+---------------------------+
| MCP server / Google |
| Cloud service / API |
+---------------------------+
How CommandLyne handles this work
CommandLyne is the enterprise AI orchestration and governance platform where teams run Gemini, Claude, OpenAI, and partner models under their own keys (BYOK). It carries the same identity principle from the user seat through to the tool call.
- Scoped tool access and permission-aware memory. Each user runs in a VM-level isolated environment with scoped tool access, permission-aware personal and team memory, and a full audit trail. See Security and Enterprise Memory (opens in a new tab).
- Per-agent tool policy for MCP. Tools are granted per agent and per workflow, so an agent reaches only the servers you name. See the Tool Policy and MCP Servers guide (opens in a new tab).
- Human sign-off checkpoints. Workflows run Preview, Approve, Adjust, Cancel, and Rollback, so a delegated action lands in a review queue before it lands in production. See Agents and Workflows (opens in a new tab).
- Browser automation under the same rules. Chrome Relay drives a real browser tab on the user's machine, inside the same policy boundary. See the Chrome Relay setup guide (opens in a new tab).
Two ways to run a delegated action
| Step | Ad-hoc agent scripts | Governed routing in CommandLyne |
|---|---|---|
| Identity | Shared service account with broad scopes | Signed-in user, delegated and scoped |
| Tool access | Every tool the runtime can reach | Per-agent MCP tool policy |
| Spend | Discovered on the monthly invoice | Token-level cost visibility per task |
| Approval | Trusted to run to completion | Preview, Approve, Adjust, Cancel, Rollback |
| Record | Logs scattered across services | Full audit trail with human sign-off |
| Model choice | One model for every step | Task-complexity, role, latency, and cost routing |
Where to start
- Inventory the identities your agents use today. List every service account, API key, and long-lived token sitting in an agent path.
- Move the first one to a delegated credential. Pick the agent that touches your most sensitive system and move it onto a user-scoped flow.
- Put checkpoints on the risky verbs. Writes, deletes, sends, and payments deserve a human sign-off step.
- Watch the cost line. Per-task token cost shows you which steps are worth a smaller model.
Try CommandLyne to run Gemini, Claude, and OpenAI agents under user-scoped credentials with per-agent tool policy, or book a demo with our team.
Try CommandLyne (opens in a new tab) | Book a demo (opens in a new tab)

