Coming from managed agents

Claude Managed Agents and the OpenAI Agents API run an agent harness in the vendor's cloud; you send input and optionally attach an environment. Cybros runs the same kind of harness on your own infrastructure and splits it in two: Nexus provides the mechanisms, your agent application (or rho) provides the policy. This page maps the vocabulary.

Concept map

Hosted platforms	Cybros	Notes
Agent (model, system prompt, tools, MCP servers, skills; versioned)	Agent member plus its declared configuration	Declared with one atomic replacement. Each new turn captures the current declaration; a running turn keeps the tools and policy it started with.
Environment (cloud sandbox or self-hosted sandbox / worker)	Runner	Always self-hosted. A runner polls its inbox over outbound HTTP exactly like a self-hosted environment worker. There is no cloud sandbox and no environment resource; environment facts are announced by the runner.
Session	Conversation plus the Runs backing its turns	The person's unit and the execution unit are separate resources.
Events (SSE stream of user/agent/tool/status events)	Inputs, turns, variants and three feeds: durable events with replay cursors, transcript deltas, progress frames	HTTP is the source of truth; the cable is a latency hint. See Events and streaming.
Send a user message	POST …/inputs with kind: direct_reply	kind: message records speech without starting a reply.
Steer or interrupt	delivery_mode: steer or steer_now; Stop	steer_now inserts a model boundary ahead of pending tools without canceling them.
Permission policy per toolset (always_allow, always_ask, auto)	Approval mode (bypass, ask, rules) plus agent-authored approval rules	Rules are globs over opaque tool input; the kernel never interprets a tool's arguments.
Scheduled deployments	Schedules	once, interval or daily; each occurrence runs in a child conversation and reports back.
Subagents, multi-agent orchestration, programmatic tool calling	delegate_task, spawn, send, wait, authored DAG steps, and the agent's code tool	Waiting, lifetime and wake are independent choices. See Orchestration.
Memory stores mounted in the sandbox	Memory documents in Nexus, read and written through tools	Separate from working files and from the opaque store.
Files and artifacts	Uploads and executor captures with resource links	Bytes are read over HTTP with range support.
Skills	Documents served by a runner or the kernel and loaded with the skill tool
Session budgets	Account cost unit, member budgets and agent allowances	Admission guards, administered per member. See Usage and cost.
Vaults for MCP or sandbox secrets	Provider credentials in Nexus; tool and MCP secrets stay with the executor	Tool processes and their credentials are executor-local state.
Webhooks	Not provided	Nexus never dials out to an application. Use the cable plus HTTP replay; see Events and streaming.
Multi-language SDKs	Ruby SDK today	The HTTP contract is primary; see SDKs and the roadmap.

What is different on purpose

The harness is yours. The loop, the sandbox protocol and the data all run in your installation. Nothing about a conversation leaves it except the model requests you configured.

Mechanism and policy are separate. A hosted platform exposes one agent object with a fixed harness. In Cybros the kernel drives the loop and persists every task, while the application declares what the tools mean, which model answers, and what needs a person's approval. Two applications can run on one Nexus with different policies.

Models are a gateway, not a vendor. Nexus resolves a model reference to one of the providers an administrator configured: an API-key provider, a subscription authorization or a local server. Fallback models, reasoning effort and prompt-format adaptation are per-model settings.

Execution is multi-environment. One run can call tools on several runners, each task recording its target. Runner effects are recorded per runner so a rewind restores each one independently.

Members, not API keys. People and agents are members of one account with stewards, workspace ownership and per-conversation access levels, rather than API keys with project scopes.

What you will miss, and what replaces it

Missing	Replacement
Cloud sandbox	The stack's in-container runner, a runner on a host or a separate runner container. Desktop runners are on the roadmap.
Webhooks	Durable event replay with cursors plus cable subscriptions.
Per-session budget	Member budgets and agent allowances at admission; an application can enforce finer limits from usage reads.
An addressable agent version number	Per-turn freezing of the declaration plus prompt document versions.
Vault	Keep tool credentials with the executor that uses them; provider keys live in Nexus as write-only settings.