Agent plan (saved procedure)

What it is

Agents are good at working out how to do something new. They are an expensive and variable way to do the same thing every week. An agent plan captures the steps once they are known. Anthropic's engineering guide makes the same distinction between agents, which choose their own steps, and workflows, which follow predefined paths; a plan is the second kind, written by the first.

Why it matters

• Cost. A plan does not spend model tokens re-planning each run.
• Consistency. The same inputs follow the same steps.
• Review. A person can read a plan, approve a version and know what will run.
• Control. Approval steps sit at fixed points rather than wherever the model decides to stop.

How it works

A plan is a list of steps. In actionlake there are seven kinds:

1. action: run one action;
2. join: combine results across systems, as in a cross-system join;
3. let: bind a value for later steps;
4. for_each: repeat over a list;
5. plan: call another plan, up to three levels deep;
6. wait: pause;
7. approval: stop until a named person decides.

Plans are saved and versioned. check validates a plan and prices it before it runs. A run is resumable: it is stored in slices and can continue on any instance. Every run carries an intent.

Example

"Every Monday at 8am, list support time logged last week in the PSA, match it to invoices in the accounting system, and send the account manager a list of unbilled hours." The agent builds it once from a written procedure, a person reads and saves it, and a schedule runs version 3 with a named key.

Common pitfalls

• Running a plan under a broad key. Give the schedule its own key, scoped to what the plan needs.
• Writes without dedupe. A retried step must not create a second invoice. See idempotency key.
• Hard-coded IDs that break when a record is renamed or moved.
• Expecting triggers. Event-driven automation is a different model; see below.

How actionlake handles it

Plans are saved, listed, versioned and drafted through the account tool, and anything an agent did once can become one: hand your agent your written procedure and it builds the plan. Schedules run a pinned plan version with a named key on a cron expression, and a leader lease makes sure scale-out never fires the same schedule twice. The plan store has 899 ready-made plans to install like a connector. There are no event triggers, webhooks or polling triggers, by design: agents and schedules start the work, and a scheduled plan can ask "what is new since the last run?". Every run goes into the ledger. See the platform page.