[ GOOGLE ADK ]
Google ADK and Gemini Enterprise Delivery
Google's Agent Development Kit is the production runtime we have shipped against since pre-launch. ADK gives us native multi-agent orchestration, a structured tool calling surface, a memory bank for persistent state, and first-class support for the A2A coordination protocol. Our Recruitment Lifecycle Platform is a working production deployment.
[ THE LEVENT POINT OF VIEW ]
We have been building on ADK since pre-launch.
When Google opened the ADK private preview, we were one of the regional teams in the room. We have shipped ADK agents in production since. The runtime is opinionated in the right places (lifecycle management, memory, A2A) and unopinionated where it should be (model choice, tool surface). It is the runtime we reach for first when the customer is on Google Cloud.
[ WHAT THIS MEANS IN PRACTICE ]
ADK's memory bank lets agents persist state across runs without bolting on a separate session store. In Smarequ, recruitment funnel state lives in Firestore with a denormalised summary map for O(1) reads, while ADK's memory layer holds the conversational context within a session. Two layers, two purposes, one runtime managing both.
Multi-agent orchestration through A2A makes our recovery pattern possible. The direct pipeline runs without LLM overhead for the happy path; when state inspection finds a partial failure, the ADK agent picks it up with full LLM orchestration. The coordination protocol carries the state forward without bespoke glue code.
Model selection within ADK is a real decision, not a default. Gemini 3 Flash for the cheap-and-fast paths (resume transcription, tool routing). Gemini 3 Pro for the consequential reasoning paths (candidate scoring, exception handling). Third-party models via the open provider interface when the workload calls for them. We design the path-to-model mapping with the cost-per-decision in front of us, not as an afterthought.
Tool integration on ADK is where most engagements gain or lose speed. ADK's native tool calling, function tools, and MCP support each have their place. We default to MCP for shared, reusable surfaces; function tools for runtime-specific glue; and built-in tools (search, code execution) where the workload obviously fits. The decision per tool drives weeks of effort over the engagement life, so we make it deliberately.
Levent is a Google Cloud Partner. Google's $750M partner programme announced in April 2026 to accelerate agentic AI delivery is something we are actively engaged on. Co-marketing on this hub and our Smarequ accelerator page is the natural surface for that partnership.
[ HOW WE DELIVER THIS ]
How we deliver this
Engineering and Build leads ADK delivery. Operate runs ADK agents in production with the AgentOps discipline they need. Enable upskills your engineers on the ADK programming model so the system survives the handover.
[ PROOF, NOT PROMISES ]
Accelerators that ship this in production today.
[ RECRUITMENT LIFECYCLE PLATFORM ]
Smarequ
An autonomous AI recruiter that screens, profiles, and orchestrates multi-round interviews. Designed to compress per-candidate processing from hours of recruiter time to minutes, with role-specific scoring the recruiter still controls.
See the accelerator →