DigitalOcean now offers a one-click Marketplace deployment for QM, Y Combinator's open-source multiplayer agent harness. The deployment provisions a Droplet, Managed Agents sandboxes, Serverless Inference, Managed PostgreSQL, and Spaces on a single bill. The post is a product availability announcement with little architectural or implementation detail.
DigitalOcean engineering blog
Loading
DigitalOcean introduces Cuttlefish, an iPhone app for its Managed Agents service that lets users follow live agent sessions, approve or deny tool calls with full command and diff context, and steer or roll back runs. The post describes sessions running in isolated DigitalOcean-hosted microVMs and a Box Mount integration that syncs Box folders into agent workspaces. Most of the content is a feature tour of a product launch.
DigitalOcean's Simulations capability, now in Public Preview, tests coding agents across full multi-turn conversations rather than single responses. The post explains the scenario, user simulator, execution sandbox, and trajectory-level judge, and shows how a baseline and a candidate agent diverge in a Python upgrade example.
DigitalOcean describes three configuration-based boundaries for its Managed Agents platform, used during a production database migration from Supabase. The layers are credential brokering through Action Gateway so the agent never sees token values, a filesystem policy profile based on NVIDIA OpenShell enforced by the runtime, and a deny-by-default network egress allowlist.
DigitalOcean Signals, now in Public Preview, analyzes production agent conversations on Managed Agents and Inference Engine and detects 25 deterministic behavioral signals across interaction, execution, and environment layers. Each signal points to the exact step that produced it, and flagged sessions can be exported as evaluation and simulation datasets. The post also reports a tau-bench study in which 82% of flagged sessions were informative, compared with 54% of random sessions.
DigitalOcean's prompt caching is now generally available on Serverless Inference, with a cache affinity router that sends repeat requests to the replica holding their prompt cache. The post describes the routing approach, the optional prompt_cache_key for isolating traffic, and self-reported hit-rate benchmarks from its own testing.
DigitalOcean warns of two critical PHP remote code execution vulnerabilities, CVE-2026-109618 and CVE-2026-109619, publicly disclosed October 21, 2026, with public proof-of-concept code available. The post lists patched PHP branches (8.5.12, 8.4.27, 8.3.36, 8.2.35), notes that PHP 8.1 and earlier receive no official patch, and reports the exposure status of each DigitalOcean product along with remediation steps.
DigitalOcean's Product Security team describes how agentic AI development has broken the traditional SDLC and how their program now embeds security through context and guardrails that agents pull in. The post walks through an SSRF mitigation pattern for webhook features and a sandbox escape that was contained by least-privilege, assume-breach design.
DigitalOcean describes the MicroVM architecture behind its Managed Agents product, which gives each session a Firecracker microVM. The post covers cell-based control planes, a fixed guest network identity per namespace, host-side credential injection, and a CAS-backed disk and userfaultfd memory restore that keeps warm starts under 200 ms.
Explains the architecture of DigitalOcean's Action Gateway for Managed Agents, which governs external tool calls separately from in-environment agent work. It covers tool resolution, search over authorized tools, per-call policy (allow, deny, ask), credential isolation from model context, chained calls via a sandboxed program, and retry semantics that distinguish safe reads from non-idempotent writes.
DigitalOcean has enhanced its Inference Router, which automatically sends each request to the model judged best for that task, based on public and internal benchmarks, cost, and latency. The router is cache-aware and exposes a preference parameter (balanced by default, with budget or quality options), and it is available in Public Preview at no extra charge beyond model usage. The post includes a basic API call example for the inference-router model.
DigitalOcean introduces Signals for triaging agent failures in production and Simulations for testing fixes through realistic, multi-turn conversations. The post explains their event pipeline, trajectory sampling, sandboxed tool execution, evaluation workflow, and regression-testing loop.
DigitalOcean introduces Agent Data Services, a managed private-preview platform for ingesting, enriching, indexing, and querying multimodal data and agent execution traces. It combines structured records and vector indexes in a governed catalog accessible through SQL and APIs, supporting RAG, agent memory, observability, and continuous evaluation.
DigitalOcean introduces Insights, a public-preview observability service that correlates metrics, logs, and traces across agents, applications, and infrastructure. Built on OpenTelemetry, it provides dashboards, PromQL queries, alerts, APIs, and access control without requiring additional instrumentation.
DigitalOcean introduces MicroVMs, Firecracker-based isolated virtual machines that boot and resume in milliseconds while preserving memory, files, and processes. The post explains how they support coding-agent sessions with checkpointing, automatic recovery, private networking, authenticated access, and usage-based pricing.
DigitalOcean introduces organization-wide resource limit management and configurable account session durations. Owners and Admins can adjust team limits across core infrastructure resources, while users can choose server-enforced sessions ranging from one hour to 30 days.
DigitalOcean introduces MicroVMs, lightweight Firecracker-based virtual machines for isolated, bursty workloads such as AI agent sessions, code sandboxes, and short-lived jobs. The private-preview service automatically pauses idle VMs and resumes them with memory, files, and processes intact, while checkpoints reduce startup overhead.
DigitalOcean introduces Agent Droplets, a bundled pricing plan for Managed Agents that combines dedicated microVM compute, hosted inference, persistent storage, and governed access to more than 16,000 tools. The post details plan tiers, usage estimates, security features, automatic pausing, and prepaid billing behavior for agent workloads.
DigitalOcean explains how its Managed Agents architecture keeps credentials outside agent contexts through execution-time brokering, scoped secrets, gateway policies, and isolated microVMs. The post details how these controls limit exfiltration, constrain agent swarms, and provide auditable, policy-driven access to connected systems.
DigitalOcean introduces Managed Agents, an integrated runtime combining isolated Firecracker microVM sandboxes, model inference, persistent data, governed tool access, and observability. The post explains its burst-based pricing, credential brokering, durable sessions, and architecture for deploying secure, stateful AI agents.