Genkit Agents API: State, Pricing, Risks, and Evaluation

Full Guide

Kingy verdict: Genkit Agents API addresses the unglamorous work that makes conversational agents usable—state, streaming, tool interruptions and frontend continuity—without forcing every team into a managed runtime. Its preview status is the decisive caveat: Google explicitly allows breaking changes in minor releases, so production adopters need version discipline and migration tests.

What the API packages

Genkit is an open-source application framework. The Agents API lets a developer define an agent on the server and use the same chat interface whether it runs in process or behind an HTTP endpoint. The object can handle a one-shot response, a streamed turn, a paused tool call and a multi-turn conversation instead of splitting those modes across unrelated abstractions.

The July 1 announcement places the preview in TypeScript and Go. Some Genkit language support extends beyond those two ecosystems, but that does not make the preview Agents API equally available everywhere. Google also warns that minor-version releases may break compatibility. Teams should pin versions and read migration notes rather than treating a routine dependency update as low risk.

State, approval and long-running work

State ownership is a genuine design choice. With a session store, the server persists messages, custom state and artifacts as snapshots while the client carries a session identifier. Without a store, the client receives the complete state and returns it on the next turn. Genkit documents in-memory and file stores for development, Firestore for shared production persistence, and a custom interface for teams with their own database, authorization or retention requirements.

The API also supports branches from an earlier snapshot, interrupts for human approval, detached tasks that can outlive a request, and delegation to specialist agents. These are useful primitives, but each adds an operational question. Who may resume a session? Can an old branch write after a newer one? What happens if approval arrives twice? How is a detached task cancelled? Which artifacts cross an agent boundary? Library support reduces plumbing; it does not answer product policy.

Pricing and architecture choice

The Genkit repository is Apache-2.0 licensed and the framework itself has no per-seat fee. Costs come from the selected model provider, hosting, Firestore or another store, telemetry and supporting cloud services. Firebase offers no-cost Spark allowances and usage-priced Blaze services, but the relevant bill depends on reads, writes, storage, functions, egress and model calls.

Google positions Genkit agents as application primitives inside a full-stack product. Its own guidance points teams toward the Agent Development Kit when multi-agent orchestration is the system’s core and a managed Agent Runtime is desired. That boundary is useful: do not choose Genkit merely because “agents” appears in the requirement.

How Kingy would evaluate it

Prototype one conversation with server-managed Firestore state and the same flow with client-managed state. Test streaming interruption, approval denial, duplicate approval, detached-task restart, branch conflict, session authorization and deletion. Pin the preview version, upgrade it in a staging branch and diff stored snapshots and protocol behavior before production.

Measure the code removed as well as the new operational surface. A successful trial should simplify frontend and tool-loop plumbing without hiding state transitions from logs or making recovery dependent on undocumented internals. Kingy reviewed the announcement, documentation, source repository and Firebase pricing; we did not build or run a Genkit agent.

Primary sources

• Google’s Genkit Agents announcement
• Genkit documentation
• Genkit agent guides
• Genkit source repository
• Firebase pricing