September 2026 Summaries

DigitalOcean describes a security architecture for autonomous agents that keeps credentials outside agents’ models and sandboxes, reducing risks from prompt injection, compromised dependencies, and unauthorized tool use. Its Managed Agents platform uses the Action Gateway to broker authenticated calls at execution time, associate access with specific actors or users, and handle authorization refreshes without exposing API keys or tokens to agents. For workloads requiring direct credentials, Harness Runtime distinguishes ordinary configuration variables, managed secrets that are injected into a sandbox, and scoped secrets that provide short-lived handles usable only with designated HTTPS destinations. Each agent and sub-agent runs in an isolated microVM with separate workspaces, histories, and credential declarations, while spawning new sessions remains subject to permissions and session limits. DigitalOcean says its policy-driven controls, including default-deny filesystem permissions, zero-egress network isolation, non-root execution, and audit trails, align with NVIDIA OpenShell’s open policy schema and aim to make multi-agent workflows more secure and auditable.

DigitalOcean has introduced Managed Agents in public preview as an integrated platform for AI agents that combines isolated execution sandboxes, model inference, persistent data, tool access, and observability in one runtime. The company argues that agents differ from conventional VM-based workloads because they operate in short compute bursts, consume tokens, wait for model or human responses, and require durable state across sessions. Managed Agents uses Firecracker microVMs for isolated, resumable sessions, an Action Gateway for governed access to more than 16,000 tools while brokering credentials outside agent environments, and active-use billing intended to charge for CPU work rather than idle time. DigitalOcean positions the service as an alternative to multi-vendor stacks, emphasizing reduced latency, centralized security, unified telemetry, support for open-weight and frontier models, and compatibility with common agent frameworks and custom OCI images. The platform targets companies building agent products, teams automating operational work, and organizations running coding agents, while planned features including Signals, evaluations, Simulations, and an Iceberg-based data platform aim to help agents learn from production outcomes; however, DigitalOcean notes that preview capabilities and roadmap plans are subject to change.

DigitalOcean has launched Managed Agents in public preview, an AI-focused platform designed to run agentic workloads without requiring teams to build and manage supporting infrastructure. The offering combines Harness Runtime, which provides durable isolated Firecracker microVM sessions with code execution, persistent storage, pause, resume, forking, parallel workflows, and observability, with Action Gateway, a managed MCP endpoint that supplies governed access to more than 16,000 tools across over 500 providers. The platform aims to address the bursty, stateful nature of agents, whose work may alternate between CPU-intensive execution and periods spent waiting for models, tools, or human approval, through per-second active CPU billing and preserved session state. DigitalOcean reports benchmark results of 886 milliseconds to runtime readiness, 3.3 seconds to a first agent response, and 305 milliseconds to resume a paused session, while acknowledging slower command round trips than Fly.io Sprites due to its authentication and audit controls. Managed Agents supports coding harnesses such as Claude Code, Codex CLI, and OpenCode, custom OCI container environments, tool permissions and approval workflows, and planned observability features through DigitalOcean Insights and Signals. Early customer Qencode reports that its support-triage agent has reduced manual triage and reporting work by an estimated four to eight hours weekly while shortening response times.

DigitalOcean has announced the general availability of Managed Databases Advanced Edition for MySQL and PostgreSQL, targeting high-growth, mission-critical applications that need greater performance, scalability, availability, and operational visibility. Following a public preview used by more than 150 customers, the company reports internal PostgreSQL benchmarks showing up to 38% higher throughput and 50% lower p99 latency, while MySQL primary-loss simulations achieved replacement-primary promotion in under three seconds on average. Advanced Edition adds rapid storage scaling, integrated proxy-based failover intended to avoid application reconnects, VPC connectivity, connection pools, multi-data-center expansion, and expanded console observability for concurrent connections and long-running queries. DigitalOcean also reduced Advanced Edition storage pricing by 46% to $0.115 per GiB per month and worked with Percona to validate the high-availability architecture. Standard Edition remains positioned for emerging projects and steadier workloads, while Advanced Edition is intended for data-intensive, high-concurrency, and enterprise production environments.

DigitalOcean CEO Paddy Srinivasan argues that AI’s “intelligence layer,” including models, agents, harnesses, data technologies, and compute infrastructure, risks becoming concentrated among a small number of closed-platform providers, potentially limiting founders’ ownership of their business differentiation. He presents open source as an alternative, citing Linux, PostgreSQL, MySQL, and Kubernetes as examples of initially less capable open technologies that became widely adopted because they offered transparency, portability, and user control. The post describes an emerging “Open Intelligence” movement driven by advancing open-weight models, open agent tools, and open data technologies, and contends that openness can also support AI safety through public inspection and evaluation rather than reliance on companies’ self-assessments. DigitalOcean highlights its long-standing open-source involvement, including repositories, managed open technologies, tutorials, and Hacktoberfest, while announcing the Open Intelligence Summit in San Francisco on October 12–13, focused on building, serving, and operating an end-to-end open AI stack.

DigitalOcean has become the infrastructure provider for Omarchy, a keyboard-focused Linux desktop built on Arch Linux and Hyprland by David Heinemeier Hansson, moving its packaging builds, PR review agents, and QA testing workloads to the platform. The company says Omarchy’s bursty, parallel agent tasks can provision a Droplet through doctl, complete a job using separately configured inference services, and automatically destroy the instance afterward. DigitalOcean is also joining the Omacom Foundation, which supports Omarchy and related open-source projects, as a Founding Corporate Patron and its agentic compute provider. The partners plan to offer a one-click Omarchy Droplet in the DigitalOcean Marketplace, allowing developers to deploy an environment matching the maintainers’ setup.