October 2026 Summaries

16 posts from DigitalOcean

Filter
Month: Year:
Post Summaries Back to Blog
DigitalOcean Simulations, available in Public Preview through its Evaluations product, assesses AI agents through complete multi-turn conversations rather than isolated responses, using a simulated user, a real agent environment with live tools, and an LLM judge that evaluates the resulting trajectory. The article illustrates the system with a Python 3.11 upgrade scenario in which an agent must retrieve current release information, let the user select a supported target version, identify breaking changes across all intervening releases, and inspect the user’s repository for affected code. Scenarios can be generated, derived from production traces, or authored in JSONL, while stopping criteria guide the simulated conversation and repeated episodes test whether an agent adapts to different user choices instead of relying on fixed answers. Teams can run baseline and candidate configurations through the DigitalOcean CLI or API, compare conversational metrics such as goal completion, knowledge retention, tool-use accuracy, and completeness, inspect tool calls and messages to diagnose failures, and pay for underlying inference and Harness Runtime usage rather than a separate simulation fee.
Oct 29, 2026 3,777 words in the original blog post.
DigitalOcean Signals, now in public preview, is an opt-in observability feature for Managed Agents on Harness Runtime and Inference Engine that analyzes production AI conversations to identify behavioral patterns that conventional infrastructure metrics may miss. Using deterministic, model-free detection across 25 signal types, it surfaces interaction issues such as user corrections, frustration, disengagement, and satisfaction; execution problems including tool failures and retry loops; and environment conditions such as rate limits, timeouts, and context overflow. Each signal links to the specific conversation step, includes supporting evidence and a confidence score, and contributes to a session-quality rating intended for triage rather than as a definitive grade. Signals requires no agent-code instrumentation, is designed to avoid the cost of using LLMs to review every conversation, and showed higher rates of useful review cases than random session selection in a controlled benchmark. Users can inspect signals through the DigitalOcean console, API, or doctl, export flagged or successful sessions as privacy-redacted datasets and simulation scenarios for evaluations, fine-tuning, or testing improvements, and manage collection through per-environment consent and deletion controls. Future plans include Action Gateway telemetry, a console interface for Inference Engine signals, scheduled exports, and an expanded signal taxonomy.
Oct 28, 2026 2,113 words in the original blog post.
DigitalOcean describes how its Managed Agents platform, in public preview, migrated huffl.ai’s production backend from Supabase to DigitalOcean while applying security boundaries aligned with NVIDIA OpenShell. The migration involved 31 database tables, more than 145,000 rows, storage buckets, and roughly 31,000 vector embeddings, with the agent provisioning infrastructure, mapping services, applying access policies, and verifying data. The platform isolates sessions in Firecracker microVMs through Harness Runtime and provides access to more than 16,000 tools through an Action Gateway that brokers credentials only when API calls are made, preventing tokens from appearing in prompts, model context, logs, or the sandbox environment. Its security model emphasizes three configuration-based controls: managed credentials that agents can use but cannot view, an OpenShell-derived filesystem policy that permits writes only to specified directories such as /workspace and /tmp, and deny-by-default network egress restricted to an explicit host allowlist. DigitalOcean states that these controls are enforced by the runtime and proxy rather than relying on agent behavior, while customers remain responsible for securing the workloads, data, and tools they connect.
Oct 26, 2026 1,158 words in the original blog post.
DigitalOcean has made prompt caching generally available for its Serverless Inference platform across supported open models, automatically discounting cached input tokens by up to 90% and reducing time to first token by up to 80%, depending on the model and prompt. The feature addresses the repeated context sent by AI agents, such as system prompts, tool definitions, files, and prior conversation history, through cache-affinity routing that directs requests to replicas already holding the relevant prompt cache rather than relying on random load balancing. DigitalOcean reports that internal October 2026 tests increased cache hit rates from roughly the mid-30% range to more than 90% for flagship models, with reported results varying by model and prompt length. The release also introduces the optional prompt_cache_key parameter to isolate cache use among teams or customers within an account, extends support to the Responses API, increases context windows to 256,000 tokens or more, and adds a dashboard showing cache hit rates, model-level performance, and estimated savings.
Oct 23, 2026 1,013 words in the original blog post.
DigitalOcean’s Product Security team argues that AI-driven and agentic development has replaced the traditional linear software development lifecycle with a fast, iterative, nonlinear process, while leaving core secure-software principles unchanged. The program focuses on secure design, security platform tooling, and offensive testing, measuring findings as security debt and balancing customer, business, and broader internet risks. Rather than relying on fixed review stages or imposing requirements on developers, the team seeks to build trust through useful security guidance, flexible remediation planning, and tools that teams voluntarily adopt. The article advocates placing mandatory security guardrails, plugins, hooks, threat models, and design patterns directly into AI agents’ workflows so agents can apply secure defaults during code generation, illustrated by an example that prevents server-side request forgery in a webhook feature. It emphasizes that established principles such as least privilege and assume breach remain essential, citing a reported sandbox escape in one of DigitalOcean’s AI products whose impact was limited by preexisting isolation, scoped credentials, and single-tenant architecture.
Oct 22, 2026 2,440 words in the original blog post.
DigitalOcean warns that two critical PHP remote code execution vulnerabilities, CVE-2026-109618 and CVE-2026-109619, disclosed on October 21, 2026, can allow attackers to run arbitrary code and potentially fully compromise servers running PHP, with public exploit code already available. All PHP versions from PHP 5 onward are considered affected, while fixes are available for PHP 8.2.35, 8.3.36, 8.4.27, and 8.5.12; users of PHP 8.1 or earlier must upgrade because those versions have no official patches. DigitalOcean advises customers to inventory every PHP installation, including CMS platforms, applications, containers, control panels, and legacy systems, patch or upgrade immediately, restart PHP-FPM or web servers, and rebuild and redeploy affected container images. The company has patched managed runtimes such as Functions and App Platform buildpacks, is patching Cloudways-managed instances, and has suspended vulnerable Marketplace 1-Click images, but customers remain responsible for PHP installed on Droplets, custom containers, Kubernetes workloads, snapshots, backups, and self-managed infrastructure. Temporary mitigations, firewalls, WAF rules, and restricting public access may reduce exposure but do not replace patching, while systems restored from older backups or images must be updated again to avoid reintroducing vulnerable PHP versions.
Oct 21, 2026 2,362 words in the original blog post.
DigitalOcean’s Managed Agents platform uses Firecracker microVMs to let agent sessions pause, resume, and fork while retaining processes, memory, files, and execution context without permanently consuming host resources. Its region-scoped cell architecture separates Kubernetes-based control-plane functions such as lifecycle management, placement, routing metadata, and PostgreSQL-backed state from hypervisor hosts that run microVMs, networking, storage, and memory services. Guests retain a fixed internal network identity through isolated Linux namespaces, while host-side Envoy proxies enforce destination policies, support private VPC connectivity, and can inject credentials for approved HTTPS integrations without exposing secrets inside the guest. Checkpoints capture disk, memory, and virtual-device state, store immutable artifacts in content-addressed object storage, and allow new branches to share base content while receiving independent writable disk overlays. Storage combines remotely addressable immutable disk layers with local sparse writable overlays, avoiding full disk downloads and remote writes for ordinary changes. During restore, userfaultfd demand-paging loads memory only when accessed, while ublk similarly serves disk content on demand, enabling a prepared template to become traffic-serving in under 200 milliseconds at the reported internal p50 under warm conditions. The design emphasizes isolated execution, fault-contained capacity cells, controlled networking, checkpoint portability, and rapid restoration as foundational capabilities for persistent, governed agent workloads.
Oct 15, 2026 5,160 words in the original blog post.
DigitalOcean’s Action Gateway is presented as a governed layer for AI agents that need to perform external actions, such as creating GitHub issues, while keeping provider credentials outside model prompts, agent environments, and model-written code. It distinguishes local agent work from credentialed external API calls, resolves the caller’s identity, eligible tools, selected connection, and policy before dispatching a request, and supports human approval for specific proposed actions rather than blanket permissions. Curated Toolbelts and MCP integration allow managed agents and external clients to discover and invoke authorized tools, while meta-tools support larger tool catalogs without exposing every schema to the model. Connections represent provider account authorizations but remain subject to Gateway policy and the provider’s own permissions, with revocation or expiration requiring reconnection. For dependent workflows, an isolated action.code environment can coordinate calls without accessing credentials, and each step remains independently subject to policy and approval checks. The system also differentiates safe retries for reads from potentially duplicate-producing writes, using idempotency protections only where providers support them and requiring verification after uncertain write outcomes.
Oct 13, 2026 2,743 words in the original blog post.
DigitalOcean has introduced enhancements to its Inference Router, a managed service that automatically selects an appropriate AI model from its catalog for each request based on task performance, cost, latency, and optional user preferences. The router supports frontier and open-source models, is initially optimized for software engineering, agentic workflows, and general reasoning, and uses benchmark and internal evaluation results to update selections as new models become available. DigitalOcean reports that its internal tests showed higher accuracy than Claude Opus 5.5 and GPT-6 Astra at roughly 22% lower cost, though results may vary by configuration and workload. Users can choose balanced, budget, or quality-oriented routing preferences, while future functionality is planned to let customers supply their own evaluation criteria. Available in public preview at no added routing charge, the service bills customers for the underlying models used and can be accessed through DigitalOcean’s console or API using inference-router as the model.
Oct 12, 2026 930 words in the original blog post.
DigitalOcean has introduced Signals and Simulations in public preview as part of its Agent Learning Services to help teams identify and test fixes for agent failures that conventional observability may miss. Signals monitors production activity from Managed Agents and the Inference Engine, grouping interactions into sessions and flagging issues across user sentiment, tool execution, and infrastructure conditions, while using low-cost triage to prioritize informative traces rather than grading every interaction. Flagged production sessions can be converted into evaluation scenarios, and Simulations uses these or user-defined scenarios to test agent or model changes through repeated, multi-turn conversations in an isolated environment with production-like configurations and tools. An LLM judge evaluates trajectories on metrics including goal completion, knowledge retention, and tool-use accuracy, allowing teams to compare baselines and candidates and distinguish reasoning problems from tool failures. The capabilities are designed to keep customer traces within a DigitalOcean team boundary, offer PII redaction, support mocked tools for actions involving data changes or costs, and enable saved scenarios to become regression tests, although preview support excludes some third-party agent frameworks.
Oct 12, 2026 1,556 words in the original blog post.
DigitalOcean has introduced Insights in public preview as a built-in observability service within its Agent Learning Services, designed to unify metrics, logs, and traces across agents, applications, and infrastructure. Built on OpenTelemetry, Insights automatically collects correlated telemetry from supported products including DigitalOcean Kubernetes, App Platform, Managed Agents, Serverless and Batch Inference, Inference Router, and Action Gateway, reducing the need for separate instrumentation or third-party integrations. It provides dashboards, signal explorers, PromQL and visual query tools, alerts, an API, and role-based access controls to help teams identify and investigate performance issues across the full request path, from agent model calls and tools to underlying pods and services. Existing DigitalOcean Monitoring metrics and alerts will continue operating until a planned transition to Insights around general availability, while the preview is free subject to usage and retention limits; DigitalOcean expects to introduce pricing at general availability and expand support for custom application telemetry and additional products.
Oct 12, 2026 1,051 words in the original blog post.
DigitalOcean has introduced Agent Data Services in private preview, a managed platform designed to centralize the data AI agents need for reliable reasoning and the execution data teams need to evaluate and improve them. The service ingests multimodal and structured sources such as documents, images, audio, video, databases, cloud storage, and agent traces, then uses configurable Data Flows to process, enrich, and vector-index the data without requiring users to manage pipeline infrastructure. It stores structured records and vectors in a governed catalog with source lineage and supports access through SQL, vector search, and APIs. Integrated with DigitalOcean services including Managed Agents, Serverless Inference, Spaces, Insights, and Agent Learning Services, it is intended to support applications such as RAG memory, media search, knowledge archives, and agent evaluation loops. The preview is offered without service-level or support commitments, and DigitalOcean notes that planned capabilities, pricing, and availability may change.
Oct 12, 2026 1,168 words in the original blog post.
DigitalOcean has launched MicroVMs in public preview, offering Firecracker-based isolated virtual machines designed for AI agents, coding platforms, sandboxes, CI/CD runners, and other on-demand execution workloads. Each MicroVM can be created from a container image, provides hardware-virtualized isolation, starts and resumes in milliseconds, preserves memory, files, and processes while paused, and incurs compute charges only while running. The service supports checkpointing to create preconfigured session environments, automatic recovery with safeguards against repeated failure loops, authenticated HTTPS access, and VPC networking for private connections to internal services. MicroVMs also underpin DigitalOcean Managed Agents, while providing developers direct control over machine lifecycle operations such as creation, pausing, resuming, checkpointing, and command execution. Pricing is usage-based at $0.022 per vCPU-hour, $0.0085 per GiB-hour of memory, $0.05 per GiB-month of storage, and $0.01 per GiB for public outbound transfer, making idle sessions less expensive because only retained storage is billed during pauses.
Oct 09, 2026 1,166 words in the original blog post.
DigitalOcean has introduced two account-management controls aimed at improving organizational oversight and user session security. Organization Owners and Admins can now use a centralized Resource Limits tab to view and adjust team limits for Droplets, Volumes, Databases, and Load Balancers within approved maximums, reducing the need to switch between teams or contact support for limit reductions. Individual users can also configure their own session duration from one hour to 30 days through their profile settings, with changes enforced server-side and logged; this setting is currently controlled by each user rather than organization administrators. Both features are available immediately without opt-in.
Oct 07, 2026 556 words in the original blog post.
DigitalOcean has introduced MicroVMs in invite-only private preview as isolated, lightweight virtual machines designed for AI agents and other bursty, short-lived workloads. Built on open-source Firecracker technology, each MicroVM runs a container image in its own hardware-virtualized environment with a separate kernel, providing stronger isolation than shared-kernel containers. MicroVMs automatically pause after a configurable idle period and resume on the next request with memory, files, and processes preserved, while checkpoints allow teams to create new instances from prewarmed environments. The service is intended for uses including agent coding workspaces, long-running workflows, reinforcement learning environments, per-tenant tools, untrusted-code sandboxes, ephemeral data stores, and CI/CD jobs. DigitalOcean says the technology already supports its Managed Agents platform, while direct MicroVM access is aimed at teams that want to build and control their own agent infrastructure or code-execution services.
Oct 01, 2026 1,053 words in the original blog post.
DigitalOcean has introduced Agent Droplets for its Managed Agents public preview, offering monthly plans intended to consolidate the infrastructure needed to run AI agents, including dedicated microVM compute, storage, hosted-model inference, persistent memory, and governed access to more than 16,000 tools. The company positions the product as a simpler alternative to combining separately billed services for agent harnesses, sandboxes, models, storage, and search from multiple providers. New users receive a $5 trial credit, while the $50-per-month Pro and $200-per-month Team plans provide 15% and 20% discounts, respectively, on eligible agent usage, with unlimited agents, environments, seats, and included enterprise security features. Once plan-supported usage is exhausted, customers can either limit spending or continue at standard rates through a prepaid Inference and Agents Balance, and active sessions can pause safely while preserving their state if funds run out. DigitalOcean says Agent Droplets support common agent frameworks and harnesses, including Claude Code, Codex CLI, OpenCode, CrewAI, and LangGraph, and compares its bundled approach with individually metered offerings from providers such as AWS, Fly.io, and Anthropic.
Oct 01, 2026 1,841 words in the original blog post.