Glossary

Every Surogate term you’ll encounter as a builder.

Term	Definition
ABAC	Attribute-Based Access Control. Policy rules evaluating user / session / tool-arg attributes. Example: “allow refund_user only if amount < 1000”.
AGENT.md	A sub-agent definition file. YAML frontmatter (name, description, tools, disallowed_tools, model, max_iterations, …) + body that becomes the spawned child’s system prompt.
Agent	A deployed service at <slug>.cloud.surogate.ai bundling a model, skills, KBs, MCP servers, persona, and channels.
API Channel	Programmatic channel. POST /v1/api/prompts with a service-account surg_sk_... token. Read results off the SSE stream at /v1/api/sessions/{id}/events — the JSON /events/poll route has no /v1/api/ alias.
Approval-required tool	A tool listed in the agent policy’s require_approval. The call is blocked and raised as a governance_gate inbox item; approving grants that exact call for one hour and one use. An explicit denial is never approvable.
Channel	A user-facing interface — Web (always on), Slack, Telegram, Website widget, API.
CLARIFY	Legacy name for ask_user_question; the response event is ask_user_question.response. Renders a tabbed widget with up to 5 questions × 4 choices each; on web/Telegram a typed reply also answers it. 30-min wait cap.
Coordinator	A session that spawns sub-agents via spawn_worker / delegate_task / spawn_task.
Event	An immutable record in a session’s log. See Event types.
Expert	A SKILL.md with type: expert. Backed by a fine-tuned model + endpoint. The base LLM delegates via consult_expert.
Goal (/goal)	Slash command that defines an outcome. The platform’s evaluator runs until satisfied / failed / blocked or 20 iterations are spent.
GRPO	Group Relative Policy Optimisation. RL training method. RL Mode: Environment / Agent / RULER.
Hub	Versioned repository browser for every artifact — skills, KBs, datasets, models, environments. Branches, commits, tags.
Inbox	Per-user queue of items the agent has explicitly raised. Five kinds: input_required, action_required, task_complete, governance_gate, progress_checkin.
MCP	Model Context Protocol. Open standard for connecting LLMs to external tools. Surogate supports stdio + HTTP transport, OAuth 2.1 PKCE.
MCP Vault	Encrypted credential storage. Secrets are injected by the MCP proxy and never reach the agent’s sandbox.
Quality flag	Per-session tag: 👍, 👎, policy.denied, harness.crash, saga.compensated, expert.override, expert.endorse. Filterable on the Sessions list.
Saga	Multi-step rollback. When enabled, sequential state-changing tool calls are checkpointed (built-in tools) or undo-declared (MCP tools); failures compensate completed steps in reverse order.
Service account	Org-scoped principal for non-interactive API clients. surg_sk_... token, valid only on /v1/api/* routes.
Session	One conversation. Append-only event log. Durable across crashes.
SFT	Supervised Fine-Tuning. The default training method.
SKILL.md	A skill file — YAML frontmatter + markdown body. type: skill (prompt-based, default) or type: expert (model-backed).
Slug	URL-safe agent name. Becomes <slug>.cloud.surogate.ai. Locked after deployment; use the Rename button to change.
SOUL.md	An agent’s persona — up to 20,000 characters. Tenant-wide, not per-user. Lives in CONFIG → SOUL.md.
Sub-agent	A declarative role preset (AGENT.md) that a coordinator applies when spawning a child session. The child inherits its parent’s agent_id and therefore its governance policy verbatim; tools / disallowed_tools are what narrow it. See How the runtime behaves.
Tool	A capability the agent can invoke. See Tool catalog.

For platform-internal terms (harness loop, sandbox runtime, lease/cursor mechanics, the events table schema), see the open-source surogates repo — they’re not relevant to building on the platform.