July 2026 Summaries

CRAB, a checkpoint-and-restore runtime developed by HKUST researchers, proposes that autonomous agent environments should be snapshotted only after recovery-relevant state changes rather than at the end of every agent turn. Using eBPF-based operating-system observation, it identifies persistent filesystem modifications, running processes, and changed memory pages, then selects the smallest safe checkpoint type, ranging from none to full state, while ignoring temporary effects that disappear before a turn ends. Evaluations reported 100% recovery correctness and checkpoint reductions of at least 70%, reaching 87% in one Terminal-Bench configuration, whereas conversation-only and filesystem-only recovery frequently failed because they omitted dependencies, services, and other runtime state. CRAB also performs checkpointing asynchronously during LLM response time to reduce latency, and the post suggests that Daytona could apply this semantics-aware policy above its existing container and VM snapshot infrastructure to make large-scale agent trajectories and reinforcement-learning rollouts more efficient.

Anchor and Daytona have collaborated to create a robust solution for running browser agents in production environments by combining their expertise in sandboxing and browser infrastructure. This partnership addresses common challenges faced by browser agents, such as security vulnerabilities, bot detection, authentication issues, and performance inefficiencies, by providing a comprehensive setup where Daytona offers isolated workspaces for agents while Anchor provides a fortified browser environment. Anchor's offerings include a hardened Chromium fork with advanced obfuscation techniques, a VPN for consistent network identity, and managed authentication services, which together ensure that agents operate like trusted human users. This integrated approach, featuring Daytona's sandboxed execution and Anchor's self-healing action layers, allows for efficient and reliable automation at scale, transforming browser agents into production-ready tools capable of handling complex tasks without frequent human intervention.

Devin sessions can now operate on Daytona sandboxes, providing each session with a dedicated Linux or Windows environment that is created from a user-defined snapshot, paused when the session is inactive, and deleted upon session termination. This setup allows Devin's tool calls to run on machines within user-controlled boundaries, which is useful when code cannot leave a specific environment or when access to internal services is needed. The integration process involves starting a Devin session, selecting a registered outpost, and using an orchestrator that manages the sandbox lifecycle, including starting, stopping, and deleting sandboxes as needed. Both Linux and Windows platforms are supported, with the integration ensuring that necessary toolchains and dependencies are incorporated into the sandbox. The setup requires Devin Enterprise and is detailed in the Devin Outposts on Daytona guide, with further documentation available through Cognition.

Brainbase’s Universal Harness API provides a single endpoint for deploying coding agents across multiple harnesses, including Claude Code, Codex, Cursor, and OpenCode, reducing the need for separate installations, configurations, and integrations. Agent definitions can include the selected harness, model, instructions, setup commands, input task, and optional capabilities such as MCP servers, secrets, and skills, while retaining a consistent request format and streaming interface. By default, each agent runs in an isolated Daytona sandbox, with Brainbase handling provisioning, filesystem and shell execution, and lifecycle management without requiring users to operate infrastructure or provide Daytona credentials. The post illustrates how the API can be used to launch a background software engineering agent that prepares a project environment, investigates and fixes failing tests, and continues work through follow-up requests, with live output streaming from the sandbox.

Daytona provides a secure and elastic infrastructure that allows users to create and manage Windows Server 2025 sandboxes, which can host applications like Microsoft Office (Word, Excel, PowerPoint) in an isolated environment. The process involves setting up a sandbox using a prebuilt Windows snapshot, installing Office with the Microsoft Office Deployment Tool, and activating it through a Microsoft 365 account. Users can programmatically drive Office applications using Daytona's computer-use API, which supports keyboard, mouse, and screenshot operations, and this automation can be observed in real-time through VNC. To optimize efficiency, users can create reusable snapshots of sandboxes with Office pre-installed, allowing for rapid deployment of Office-ready environments for various use cases, such as document automation and testing. Daytona's infrastructure facilitates the secure execution of AI-generated code, with tools available for managing these processes via SDKs, CLI, and API.

OSWorld serves as a prominent benchmark for evaluating computer-use agents by giving them natural language instructions to perform tasks on live desktops using screenshots, a mouse, and a keyboard, with a script inspecting and scoring their attempts. While the benchmark includes tasks for both Ubuntu and Windows operating systems, the latter requires setting up a virtual environment manually, as there is no pre-packaged image available. The piece details an experiment where Windows tasks were executed using Daytona Windows sandboxes, focusing on the performance of Claude Opus 4.6 and GPT-5.4 agents. These agents were tasked with activities ranging from single-app tasks in Microsoft Office to complex multi-app workflows, with results showing varied success rates across different categories. The document highlights the process of setting up the environment, running tasks, and the modifications made to evaluators to address scoring issues, further providing a comprehensive guide for replicating the experiment.

Kubernetes excels at managing and scaling replicas of known services, but its efficiency diminishes when dealing with diverse, short-lived AI workloads that require distinct execution environments. In a comparative evaluation, the Daytona system demonstrated significantly faster startup times, maintaining a median time-to-first-command close to one second for various working sets, while Kubernetes took between 4.2 and 4.4 seconds. Despite its predictability, Kubernetes incurred higher operational costs for launching multiple distinct environments. The results suggest that Kubernetes is best suited for stable, long-term services, whereas Daytona's prepared sandboxes are more effective for rapidly launching diverse, transient environments. This evaluation highlights the importance of choosing the right platform based on workload requirements, emphasizing the need for tailored solutions in managing diverse AI workloads.

Daytona has introduced a new feature that allows sandboxes to operate as true microVMs, providing machine-level isolation with each having its own guest operating system, kernel, and virtualized hardware. This isolation enhances security by making it safe to run untrusted or AI-generated code. The system supports features like "Pause" and "Fork," where "Pause" freezes the entire machine state, allowing it to resume in under a second without consuming compute resources, while "Fork" creates multiple identical VM copies from a single point in time, enabling parallel workflows and deterministic comparisons. These functionalities are available to all Daytona organizations, aimed at improving infrastructure for running AI-generated code efficiently. Juraj Stefanič, a Technical Content Engineer at Daytona, plays a crucial role in helping developers understand and adopt these advanced infrastructure capabilities through technical content and documentation.

Domain Firewall offers domain-level egress control for sandboxes, enabling precise management of outbound network requests by allowing only specified destinations and blocking others, thus preventing unauthorized data exfiltration and access to restricted services. This feature is particularly important for enterprises running untrusted or AI-generated code, as it ensures that sandboxed environments can only communicate with a predefined set of domains, addressing security concerns during code execution. The mechanism operates on a deny-by-default policy, where any domain not explicitly allowed is automatically blocked, and is enforced at the sandbox level to accommodate various network policies across different workloads. Available to all Daytona organizations, Domain Firewall helps align the network reach of code with its actual needs, thereby facilitating secure and compliant sandbox operations.

Secrets Manager is a tool designed for securely managing credentials needed for AI-generated code execution within Daytona sandboxes, ensuring that sensitive information like API keys and access tokens remain hidden from the agents using them. By injecting secrets at runtime and masking their real values, Secrets Manager allows agents to authenticate to external systems without direct access to the underlying credentials, thereby reducing the risk of exposure when running untrusted or AI-generated code. This system separates access from visibility, with secrets scoped and managed per organization, available only to the specific sandboxes that require them. The tool is available to all Daytona organizations, and users can utilize the provided documentation to implement secret management within their workflows. Juraj Štefanić, a Technical Content Engineer at Daytona, plays a crucial role in helping developers adopt this infrastructure, creating content that bridges the gap between sandboxing technology and practical developer workflows.

Daytona's Windows Sandboxes offer real Windows environments with instant virtual machine (VM) provisioning, enabling automation and testing to run against genuine Windows software rather than Linux approximations. This ensures that desktop applications, enterprise tools, and automation targets that require Windows operate as they would in production. The sandboxes provide a fully isolated VM running the native Windows operating system, complete with registry and APIs, and they can be provisioned in seconds, unlike traditional VMs that take minutes to boot. These sandboxes are particularly useful for desktop application testing, robotic process automation (RPA) workflows, and Windows-specific automation, as they allow users to immediately run regression suites, automate business processes, and drive workflows requiring Windows-specific features. Daytona offers snapshots for creating sandboxes in various sizes to meet different computational needs, ensuring flexibility and efficiency for all users.

GPU Sandboxes offer dedicated GPU access to ensure consistent performance and hardware-level isolation for GPU-bound workloads, such as reinforcement learning, model training, and inference. By using VFIO passthrough, each sandbox receives a direct connection to a physical GPU, eliminating the performance variability and isolation issues associated with shared or virtualized GPU resources. Supported GPU models include NVIDIA H100, H200, RTX Pro 6000, RTX 4090, and RTX 5090. This setup allows for predictable and uncontested GPU access, making it ideal for applications requiring reliable and direct hardware interaction. GPU Sandboxes are available to Daytona organizations, who can access detailed documentation to facilitate their implementation.