Skip to content

Agent primitives

The durable services that let agents work across time, tools, and teams.

Dexto Cloud is built from a small set of agent-compatible primitives. Each one solves a different part of the job: keeping work alive, giving it compute, connecting it to the systems it needs, and making the output reviewable.

The Dexto Cloud control plane keeps the durable record for agents, sessions, runs, approvals, and artifacts. Cloudflare Durable Objects coordinate live session state, active runs, reconnects, and short-lived locks; Workflows coordinate asynchronous work and retries.

You do not provision these services yourself. They are the runtime layer behind a Dexto Cloud agent.

PrimitiveWhat it ownsWhat it does not promise
Agent and session stateInstructions, conversation history, run records, and configured toolsEvery process or unsaved workspace file surviving a runtime failure
Live coordinationActive-run state, reconnect handling, locks, and bounded runtime snapshotsA globally queryable product database
Workflow orchestrationAsynchronous steps, waits, retries, and completion handlingPermission to repeat a consequential action without policy and approval
Artifacts and approvalsDurable deliverables and explicit review requestsAutomatic sign-off or publication

Promote work that matters to artifacts. Workspace files and durable product outputs have intentionally different lifecycles.

When a task needs code, files, or a browser, an agent can work in a cloud computer. Dexto Cloud keeps the task and its durable state in the control plane while the computer provides the temporary execution environment.

This separation lets a team review the work without treating an execution workspace as the source of truth. It also means a sandbox does not override the permissions of a connected system: access to production data still depends on the connection and policy you give the agent.

Dexto Cloud routes LLMs, image generation, voice and audio, video, open-source LLMs, and managed compute through one services catalog. An LLM can be Dexto Cloud-routed or use an account-owned provider connection where supported.

The catalog exposes the capability, modalities, pricing model, and availability of each entry so a team can choose a model for the job rather than hard-coding one provider into every workflow.

Connections give agents access to the services where work already happens: source control, knowledge, CRM, analytics, communication, and more. Credentials remain inside platform-controlled execution paths; an agent receives the permitted tool surface, not a raw provider secret.

You can also add a custom remote MCP server. Its available tools are discovered after connection and remain subject to the connected account and the context in which an agent runs.

Projects organize related conversations, sources, and instructions. Memory keeps useful context available to the right work. Approvals give a person a clear moment to inspect and decide before a configured consequential action continues.

Together, those services make a run understandable after the fact: what it was asked to do, what it used, what it produced, and which decisions required a human.

Dexto Cloud can deliver work through the app itself, Apps, Sites, an OpenAI-compatible API, or MCP. The interface changes; the underlying agent, context, and connected capability boundaries stay explicit.