Search
Tech

Build Long-running AI agents that pause, resume, and never lose context with ADK

"Enterprise-grade AI agents are finally within reach, thanks to a new architectural framework that enables long-running workflows to pause, resume and maintain context across days or weeks. By leveraging durable state machines and persistent session storage, developers can build resilient systems that 'sleep' during downtime and wake up with high reasoning accuracy. The key lies in event-driven webhooks and multi-agent delegation, which together ensure seamless transitions and uninterrupted workflows."

Chris L (AI-assisted) 1 min read EN

Based on reporting from Source.

Google has released the Agent Development Kit (ADK), a framework for building long-running AI agents that can pause, resume, and maintain context across days or weeks. ADK enables developers to create resilient systems that 'sleep' during downtime and wake up with high reasoning accuracy. The framework uses durable state machines, persistent session storage, event-driven webhooks, and multi-agent delegation to ensure seamless transitions and uninterrupted workflows.

Overview

ADK is designed to address the limitations of stateless chatbots, which struggle to manage long-running enterprise workflows that span days or weeks. The framework provides a new architectural approach that separates production agents from demo chatbots.

What ADK Does

ADK enables developers to build long-running AI agents that can manage complex workflows, including HR onboarding, invoice dispute resolution, and sales prospecting sequences. The framework provides a set of architectural shifts, including:

  • Durable memory schemas instead of dumping raw JSON into a vector database
  • Event-driven dormancy gates instead of active polling or blocked threads
  • Multi-agent delegation instead of monolithic single-agent prompts

How ADK Works

ADK uses a combination of durable state machines, persistent session storage, event-driven webhooks, and multi-agent delegation to ensure seamless transitions and uninterrupted workflows. The framework provides a set of tools and APIs for building long-running agents, including:

  • Persistent session storage using SQLite or Cloud SQL
  • Event-driven webhooks for handling external events
  • Multi-agent delegation for specialized tasks
  • Evaluation frameworks for testing and validating agent behavior

Tradeoffs

ADK requires a fundamental shift in architectural approach, which can be challenging for developers who are used to building stateless chatbots. However, the benefits of ADK, including improved reasoning accuracy and reduced token costs, make it a compelling choice for building long-running AI agents.

When to Use ADK

ADK is suitable for building long-running AI agents that manage complex workflows, including HR onboarding, invoice dispute resolution, and sales prospecting sequences. The framework is particularly useful for workflows that involve human-in-the-loop pauses, cross-system handoffs, or multi-day timelines.

Bottom Line

ADK provides a powerful framework for building long-running AI agents that can pause, resume, and maintain context across days or weeks. The framework's durable state machines, persistent session storage, event-driven webhooks, and multi-agent delegation make it an ideal choice for building complex workflows that require high reasoning accuracy and reduced token costs.

Sources Used

  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [Source Name]
  • [

Sources 1

  1. 01 Source https://developers.googleblog.com/build-long-running-ai-agents-that-pause-resume-and-never-lose-context-with-adk/
Similar Articles

More articles like this

Tech 1 min

Nuclear Diamond Batteries, Inc. (OTC: NDBI) Receives Second U.S. Patent Allowance Advancing Commercialization of Ultra-Long-Life Nuclear Diamond Battery Platform

Nuclear Diamond Batteries’ second U.S. patent approval accelerates its betavoltaic power cells—carbon-14 diamond layers that harvest electron flux for 28,000-year half-life micro-power—toward commercial pilots in IoT sensors and pacemakers, slashing battery-replacement logistics in remote and implantable devices. The allowance covers stacked, hermetically sealed modules that boost output to milliwatt scale while meeting ISO 13485 medical-grade standards.
Tech 1 min

OUTCOME2030 Brings Industry Leaders Together to Examine How AI, Private Capital, and Employee Benefits Support Working-Class Communities

Capital flows and technological innovation converge on working-class communities as industry leaders gather to dissect the intersection of private investment, AI-driven workforce solutions, and employee benefits in a bid to redefine economic mobility. The conference, OUTCOME2030, will scrutinize the impact of capital allocation, AI-powered tools, and benefits packages on low-income households. Key stakeholders will convene in Washington, DC, to examine the mechanisms driving economic mobility.
Tech 1 min

Eason Technology Limited Issues Statement Regarding Unusual Market Action

Unprecedented trading halts on the NYSE American exchange prompt Eason Technology Limited to clarify the circumstances behind a sudden, unexplained spike in stock volatility, citing "ongoing internal review" and assuring investors of "no material adverse change" in the company's financial position. The statement comes as the firm's shares experience a 10% fluctuation in a single trading session. Regulatory scrutiny is now focused on the incident.
Tech 1 min

Safe Air Technology Partners with Milton Street Capital and EA Advisors to Support Scaling of Mission-Critical Data Center Infrastructure

With a $450M majority recapitalization led by Milton Street Capital, Safe Air Technology is poised to scale its custom, high-efficiency HVAC systems—specifically its 2N+1 redundant chillers and direct-to-chip liquid cooling loops—across hyperscale data centers, addressing the thermal bottleneck of AI workloads before NVIDIA’s next-gen Blackwell GPUs hit full deployment.
Tech 1 min

Hubble Connected Partners with Kodak to Launch New Family Technology Ecosystem Across Baby, Kids, and Pet Categories

A new family technology ecosystem emerges, integrating Kodak's legacy imaging capabilities with Hubble Connected's smart communication platform to create a comprehensive care-focused network for baby, kids, and pet monitoring, leveraging a suite of connected devices and AI-driven insights to streamline family life. The partnership combines 130 years of Kodak's imaging expertise with Hubble Connected's IoT platform, now supporting over 1 million families worldwide. This strategic collaboration sets a new standard for family-centric smart home solutions.
Tech 1 min

Trellis Brings the Largest State Trial Court Dataset in the U.S. to Claude -- and Anyone Can Try it for Free

A landmark dataset of 45-state trial court intelligence, encompassing rulings, verdicts, and judge analytics, has been integrated into Claude, a large language model, marking a significant expansion of its knowledge base. This vast repository, courtesy of Trellis, the largest legal research platform, now enables users to tap into a unified, nationwide archive of state trial court decisions. The dataset is available for free, allowing researchers and developers to explore its potential.