Ananke.Federation.Google 0.8.1

Ananke.Federation.Google

Platform update (April 2026): Google has rebranded Vertex AI as Gemini Enterprise Agent Platform. All Vertex AI services and roadmap evolutions are now delivered exclusively through Agent Platform. The VertexAI* class names in this package are preserved for backwards compatibility; new platform capabilities are documented below under New platform capabilities.

Gemini Enterprise Agent Platform adapter for Ananke Federation. Deploy Ananke workflow manifests to Agent Runtime, monitor running agents via Agent Observability, and route executions through Gemini-hosted models.

What this package provides

Authentication

This adapter uses Application Default Credentials (ADC). Provide your Google Cloud project ID and region:

var credentials = new VertexAICredentialProvider(
    project: "my-gcp-project",
    location: "us-central1");

In local development, run gcloud auth application-default login. In CI/Cloud Run, the attached service account is picked up automatically.

The service account requires these IAM roles:

Quick start

Deploy a workflow

var credentials = new VertexAICredentialProvider("my-gcp-project", "us-central1");
var registry    = new InMemoryDeploymentRegistry();
var deployer    = new VertexAIDeployer(credentials, registry);

DeploymentRecord record = await deployer.DeployAsync(manifest, toolKit, new DeployOptions
{
    Platform = "vertex-ai"   // or "gemini-agent-platform" — both accepted
});
// record.Status == DeploymentStatus.Active
// record.PlatformResourceId == "projects/my-gcp-project/locations/us-central1/agents/<id>"

Monitor a deployment

var monitor = new VertexAIRemoteCellMonitor(
    project: "my-gcp-project",
    options: new RemoteCellMonitorOptions
    {
        LookBackMinutes   = 30,
        ErrorRateThreshold = 0.05,   // unhealthy above 5% errors
        LatencyThresholdMs = 5_000   // unhealthy above 5 s avg latency
    });

RemoteCellHealth  health  = await monitor.GetHealthAsync(record.DeploymentId);
RemoteCellMetrics metrics = await monitor.GetMetricsAsync(record.DeploymentId);

Tool translation

VertexAIToolSchemaTranslator groups tools by execution mode:

Local tools throw InvalidOperationException — remote cells cannot call back to the supervisor process (supervisor-only hybrid constraint).

Model mapping

VertexAIModelMapper maps any Ananke model reference to its Gemini equivalent. Google models pass through unchanged; OpenAI and Anthropic models are mapped to the nearest Gemini equivalent.

Current default mappings (updated for Agent Platform GA):

Pinned 2.5-class model strings (e.g. gemini-2.5-pro) still pass through unchanged if you have an explicit reason to target them.

Monitoring

VertexAIRemoteCellMonitor queries two Google Cloud APIs:

• Cloud Trace v2 (cloudtrace.googleapis.com/v2/projects/{p}/traces) for per-invocation trace records. Each span is checked for error status and latency.
• Cloud Monitoring v3 (monitoring.googleapis.com/v3/projects/{p}/timeSeries) for aggregated execution, token, tool-call, and error-count metrics over the configured LookBackMinutes window.

Health is computed as:

IsHealthy = errorRate ≤ ErrorRateThreshold AND avgLatencyMs ≤ LatencyThresholdMs

When the Observability API is unavailable (e.g. network error, missing IAM role), GetHealthAsync returns IsHealthy = false and GetMetricsAsync returns zero values — both degrade gracefully without throwing.

Platform adapter status

New platform capabilities (Agent Platform GA)

The following Agent Platform features are newly available and represent planned or candidate work for this adapter:

Agent Memory Bank

Agent Memory Bank provides persistent, long-term memory with Memory Profiles. This maps naturally to Ananke's IConversationMemory and InMemoryEmpiricalMemory abstractions — a GoogleMemoryBankAdapter could back both with durable, cross-session storage.

Agent Sessions

Agent Sessions with Custom Session IDs allow Ananke checkpoint/resume IDs to be mapped directly to Agent Platform session records, enabling end-to-end traceability from ICheckpointStore to the Google Cloud audit trail.

Agent Identity & Agent Registry

Agent Identity assigns a cryptographic ID to every deployed agent. Combined with the Agent Registry (central library of approved agents, tools, and skills), this provides the governance layer that complements Ananke's IDivisionApprovalGate and InMemoryCapabilityMap.

Agent Gateway

Agent Gateway acts as the control point for all agent traffic, enforcing Model Armor protections against prompt injection and data leakage. Ananke Federation deployments should route through Agent Gateway to inherit these enterprise security guardrails.

Agent Simulation & Evaluation

Agent Simulation and Agent Evaluation provide automated quality scoring and multi-turn autoraters. These are complementary to Ananke's OfflineLearner / InMemoryEmpiricalMemory — Google scores task success and safety; Ananke scores domain outcomes.

Agent Sandbox

Agent Sandbox provides a hardened environment for model-generated code execution and browser-based automation. Relevant for Ananke workflows that include SkillCatalogDemo-style tool invocations running on deployed cells.

Bidirectional Streaming

Agent Platform supports WebSocket-based bidirectional streaming for live audio and video. This aligns with Ananke's IStreamingAgentModel and opens a path for real-time multimodal AgentJob scenarios on deployed cells.

Migration notes (Vertex AI → Agent Platform)

No code changes are required in consuming applications — the VertexAI* class names are stable. Update model strings in secrets.json / configuration if you have pinned gemini-2.5-* identifiers.

What this package provides

Authentication

This adapter uses Application Default Credentials (ADC). Provide your Google Cloud project ID and region:

var credentials = new VertexAICredentialProvider(
    project: "my-gcp-project",
    location: "us-central1");

In local development, run gcloud auth application-default login. In CI/Cloud Run, the attached service account is picked up automatically.

Quick start

var credentials = new VertexAICredentialProvider("my-gcp-project", "us-central1");
var registry = new InMemoryDeploymentRegistry();

var deployer = new VertexAIDeployer(credentials, registry);
DeploymentRecord record = await deployer.DeployAsync(manifest, toolKit, new DeployOptions
{
    DeploymentId = "my-agent-v1",
    Platform = "gemini-agent-platform"
});

Tool translation

VertexAIToolSchemaTranslator groups tools by execution mode:

Local tools throw InvalidOperationException — remote cells cannot call back to the supervisor process (supervisor-only hybrid constraint).

Model mapping

VertexAIModelMapper maps any Ananke model reference to its Gemini equivalent. Google models pass through unchanged; OpenAI and Anthropic models are mapped to the nearest Gemini equivalent.

Current default mappings (updated for Agent Platform GA):

Third-party models (Claude Opus, Sonnet, Haiku) remain accessible directly through Model Garden without mapping.

Monitoring

VertexAIRemoteCellMonitor integrates with Agent Observability to surface execution traces, latency, token usage, and error rates. Visual trace inspection is available in the Agent Platform console.

Platform adapter status

New platform capabilities (Agent Platform GA)

The following Agent Platform features are newly available and represent planned or candidate work for this adapter:

Agent Runtime (unblocks DeployAsync / TeardownAsync)

Agent Runtime is now GA with sub-second cold starts and agent provisioning in seconds. It supports long-running agents that persist state for multiple days — enabling Ananke workflows that require extended execution windows (e.g. multi-step research, sales prospecting sequences). See Agent Runtime docs.

Agent Memory Bank

Agent Memory Bank provides persistent, long-term memory with Memory Profiles. This maps naturally to Ananke's IConversationMemory and InMemoryEmpiricalMemory abstractions — a GoogleMemoryBankAdapter could back both with durable, cross-session storage.

Agent Sessions

Agent Sessions with Custom Session IDs allow Ananke checkpoint/resume IDs to be mapped directly to Agent Platform session records, enabling end-to-end traceability from ICheckpointStore to the Google Cloud audit trail.

Agent Identity & Agent Registry

Agent Identity assigns a cryptographic ID to every deployed agent. Combined with the Agent Registry (central library of approved agents, tools, and skills), this provides the governance layer that complements Ananke's IDivisionApprovalGate and InMemoryCapabilityMap.

Agent Gateway

Agent Gateway acts as the control point for all agent traffic, enforcing Model Armor protections against prompt injection and data leakage. Ananke Federation deployments should route through Agent Gateway to inherit these enterprise security guardrails.

Agent Simulation & Evaluation

Agent Simulation and Agent Evaluation provide automated quality scoring and multi-turn autoraters. These are complementary to Ananke's OfflineLearner / InMemoryEmpiricalMemory — Google scores task success and safety; Ananke scores domain outcomes.

Agent Sandbox

Agent Sandbox provides a hardened environment for model-generated code execution and browser-based automation. Relevant for Ananke workflows that include SkillCatalogDemo-style tool invocations running on deployed cells.

Bidirectional Streaming

Agent Platform now supports WebSocket-based bidirectional streaming for live audio and video. This aligns with Ananke's IStreamingAgentModel and opens a path for real-time multimodal AgentJob scenarios on deployed cells.

Migration notes (Vertex AI → Agent Platform)

No code changes are required in consuming applications — the VertexAI* class names are stable. Update model strings in secrets.json / configuration if you have pinned gemini-2.5-* identifiers.