External Runtimes

Agents issues managed accounts to external runtimes.

That is the canonical umbrella term for the thing outside Craft that uses a managed account token.

Typical external runtimes include:

• agents
• orchestrators
• workers
• scripts

Why Use An Umbrella Term

Not every integration looks the same.

Some teams run:

• one cron-driven worker
• one agent runtime that fetches data and reasons directly
• one orchestrator that coordinates tools and submits approvals
• a hybrid of deterministic scripts plus an LLM step

Agents should describe all of those cleanly without pretending they are identical.

So the product language is:

• Agents = the product and trust boundary
• managed account = the machine identity
• external runtime = the thing using that identity

Trust Rule

Use this as the calm mental model:

• operators set the boundary
• external runtimes work inside it
• Agents enforces the rules

That means operators decide:

• scopes
• optional approvals
• write boundaries
• who receives operational visibility

And the external runtime does the execution work inside those limits.

When A Worker Is Helpful

A worker is just one kind of external runtime.

It is often the best pattern when you want:

• deterministic fetch and preparation
• scheduled cron-style execution
• retries and backoff without LLM token spend
• repeatable input datasets for audits or review

That is why the docs still include worker-first examples.

When A Direct Agent Orchestrator Is Enough

A separate worker is not mandatory.

A direct agent or orchestrator can use the managed account token itself when:

• the workflow is simple
• the data inputs are bounded
• the runtime already handles scheduling and tool calls well
• reproducibility pressure is lower

In that model, the agent or orchestrator can:

• fetch data from Agents
• reason over it
• call other tools
• submit governed approval requests

What Agents Does Not Do

Agents does not run the external runtime for you.

It does not:

• execute model reasoning
• orchestrate external tool calls
• run arbitrary shell commands as production behavior

Instead, it provides:

• managed accounts
• scoped API access
• approvals for sensitive actions
• diagnostics and runtime visibility
• a governed boundary between Craft and the external runtime

Rule Of Thumb

If you need the simplest first bootstrap:

• start with a worker

If you already have a disciplined orchestrator or agent runtime:

• it can use the managed account directly

Both patterns fit the product.

See Also

• First Worker
• Execution Model
• Job Guides