A closer look at Agents CLI
Summary
In this Google Cloud Next session, AI Product Manager Shubham Saboo joins host Jason Davenport to walk through three major developer launches designed to move AI agents from experimental demos to production-grade systems. The conversation covers the Agent CLI, ADK 2.0, and the concept of Skills, as well as supporting capabilities like long-running agents, ambient agents, pause-and-resume, and Agent Garden.
The Agent CLI is presented as the new entry point for the entire agent development lifecycle. It is a command-line tool combining skills and commands that lets developers scaffold, build, deploy, evaluate, and add observability to agents across the Gemini Enterprise Agent Platform. Crucially, the CLI integrates with any coding agent — Gemini CLI, Claude Code, Codex, or others — via a global install, giving those agents full contextual knowledge of ADK and the Agent Platform without manual configuration.
The concept of Skills is explained as a cross-industry standard (not Google-specific) that replaces the practice of stuffing all agent instructions into a single monolithic system prompt. Skills give agents modular capabilities they can select and invoke at runtime, making agents more adaptable and preventing context window overload. The Agent CLI itself is built on a library of skills covering platform knowledge, scaffolding, deployment, evaluation, and observability.
ADK 2.0 introduces graph-based workflows, which bring determinism to inherently non-deterministic LLM agents. Developers can now define precise routing logic and task execution sequences, making the approach well-suited for regulated industries like financial services and insurance — scenarios where some steps must execute in a specific, predictable order even while the surrounding system remains agentic.
Three production-readiness features are highlighted. Long-running agents can now maintain state for up to seven days, enabling agents to persist context across multi-day projects rather than restarting fresh each session. Ambient agents can trigger based on external events — incoming messages, schedules, or other signals — rather than only responding to direct prompts. Pause-and-resume agents address the real-world challenges of network interruptions and human-in-the-loop approvals: a single parameter enables an agent to save its state, pause indefinitely, and pick up exactly where it stopped — whether the pause lasts minutes or days.
Agent Garden is introduced as a library of pre-built multi-agent templates and patterns available on GitHub, covering sequential, loop, and parallel agent architectures, as well as more advanced patterns like human-in-the-loop, coordinator-dispatcher, and iterative refinement. The session closes with Saboo emphasizing that combining all these capabilities — flexible graph workflows, ambient and resumable agents, and long-running state — marks a meaningful shift from agents that run for minutes to agents that can reliably operate for days and solve real production problems.
Key Takeaways
- •Install the Agent CLI globally and it will be picked up by any coding agent (Gemini CLI, Claude Code, Codex), giving it full ADK and platform context automatically.
- •Replace monolithic system prompts with Skills — modular capabilities agents can select at runtime — to reduce context overload and improve agent reliability.
- •Use ADK 2.0 graph-based workflows when your process has deterministic routing steps; reserve LLM non-determinism for only the parts that need it.
- •Long-running agents can now maintain state for up to 7 days, enabling multi-day project continuity without restarting from scratch.
- •Enable pause-and-resume with a single parameter in ADK to handle network drops, coffee breaks, and human-in-the-loop approval delays without losing agent state.
- •Use ambient agents to trigger workflows from external events (messages, schedules) rather than requiring a direct user prompt each time.
- •Start with Agent Garden's pre-built templates (sequential, loop, parallel) before building custom multi-agent architectures from scratch.
- •Ask the Agent CLI what tools and eval options exist rather than reading docs — it has full platform context and can walk you through setup interactively.
- •For regulated industries like insurance claims or financial services, graph-based routing provides the reliability and auditability that pure LLM agents cannot.
- •The fundamental multi-agent patterns (sequential, loop, parallel) compose into all advanced patterns — master these first before tackling coordinator-dispatcher or iterative refinement architectures.
Transcript
Hey, what's up friends and family? Back here with Jason Davenport with Google Cloud and I am joined in this next session with Shubham Saboo, an AI product manager here at Google Cloud. And Shubham, super excited to have you on the show and to talk about all of the cool things that we're doing here at Cloud Next. So, let's let's get into it. Uh what is the most exciting things that we're doing for developers here that we've launched here today at Google Cloud? Yeah. Thank you for having me, Jason...
