Skip to main content
ECP defines a Context as the root execution object: a portable artifact that describes a complete execution environment for one or more AI agents.

​
Context

A Context contains:
  • Inputs — parameters that make the Context reusable (e.g. shopifyStoreId, jiraProject)
  • Outputs — declared structured results (e.g. from a schema)
  • Schemas — JSON Schema-like definitions referenced by executors and outputs
  • Triggers — events that start a run (schedule, webhook, tool-event)
  • Orchestration — entrypoint, strategy, and coordination defaults
  • Orchestrator — the top-level coordinating object (optional; legacy Contexts use a flat executors list)
Contexts are portable: version them, share them, and run them on any ECP-compliant runtime.

​
Orchestrator and executors

  • Orchestrator — The entry point of execution. It may contain executors and nested orchestrators, define orchestration strategy, and coordinate delegation and aggregation.
  • Executors — Execution roles (agents, tools, or human reviewers) that run within the orchestration. Each has its own protocols, mounts, and policies.
Orchestrators and executors share a common pattern for inputs, outputs, and schema references. Orchestrators add coordination-specific properties (child executors, delegation semantics, merge responsibilities).

​
Orchestration strategies

StrategyDescription
singleRun only the entrypoint; no delegation.
sequentialRun children in a fixed order; outputs can feed the next.
delegateOrchestrator plans, delegates to specialists, then merges results. Most common for complex flows.
swarmRun children in parallel; aggregate or merge afterward.

​
Mounts

Mounts define how data is loaded from external systems (e.g. via MCP tools). They run in three stages:
StagePurpose
SeedMetadata and references (e.g. Ref objects with id, title, snippet).
FocusA subset of objects selected during orchestration (bounded).
DeepFull documents or large payloads (used sparingly).
This staged hydration keeps context windows and cost under control.

​
Policies

Policies define object-scoped security in the Context manifest (per executor):
  • Tool access — Default deny; allow only explicitly listed tools (e.g. jira:issues.search).
  • Budgets — e.g. maxToolCalls, maxRuntimeSeconds.
  • Write controls — forbid, propose-only, or execute-if-allowed; optional requireApprovalFor list.
The host may still apply a separate layer: ecp.config.yaml with a top-level security mirror (allow-listed model providers, MCP server names, plugin kinds, secret providers, …). See Security and the system configuration section in the Schema reference.

​
Protocols

Executors declare which protocols they use, for example:
  • agentOrchestration — A2A (agent-to-agent)
  • toolInvocation — MCP
The runtime uses these to wire tool calls and delegation.

​
Triggers

Triggers start a Context run. Supported types include:
  • schedule — Cron-style (e.g. 0 8 * * MON with timezone)
  • webhook — HTTP-triggered
  • tool-event — Emitted by an integration

​
Learn more