AWS AgentCore action
Invokes an agent you host in AWS—a Bedrock AgentCore agent runtime. In the action picker it is AWS AgentCore, under AI Actions. The agent's progress streams into the run as it works, and it can pause to ask a person a question.
Fields
| Field | Required | Description |
|---|---|---|
| AWS Account | Yes | The AWS connection whose runtimes this step can reach, in that account's region. Pick it first—everything below is drawn from it. |
| Agent | Yes | The hosted agent runtime to invoke, chosen from the runtimes the selected account holds in its region. |
| Input | Yes | What to send the agent—a Prompt, written as text, or JSON, a structured payload. Both accept variables, resolved before the request is sent. |
| Long-term Memory Scope | No | How much the agent recalls between runs—see below. |
| Human in the Loop | No | Opt-in. Set it to let the agent pause and ask a person before finishing, and choose the Providers the question is pushed to. Left unset, the agent must finish on its own. |
AWS Account comes first for a reason: the list of agents depends on it. Leaving it blank means the deployment's own AWS credentials, and only where the deployment permits that—where it does not, the step cannot run until this field names an account.
Long-term Memory Scope
This decides which prior runs an agent can recall. It has no effect unless long-term memory is enabled on the agent's own memory resource in AWS; where it is, facts and summaries accumulate per scope.
| Scope | What shares a memory |
|---|---|
| Per Node | This step alone. Two AgentCore steps in the same automation stay separate. |
| Per Workflow | Every AgentCore step in this automation, across all its runs. The default, and the one that matches "this automation remembers what it learned". |
| Per Site Agent | The same agent, wherever it is invoked. Right for an agent that acts as a standing assistant rather than belonging to one automation. |
| Per Site | Every AgentCore step in every automation. The broadest scope, and the easiest way for unrelated automations to leak context into each other—pick it only when sharing is the point. |
| Ephemeral | Nothing. The run leaves no trace a later run can recall. Use it for test runs and for content that should not persist. |
Short-term recall within a single run, including the conversation around a human-in-the-loop pause, works the same whichever scope you pick.
Pausing for a person
With Human in the Loop set, the agent can stop to ask a question and the step waits for the answer before the agent resumes. Pick the Providers the question is pushed to—Slack, Microsoft Teams, or both—and turn on Only Notify on Questions to push the pauses alone rather than the running progress. Every request is also answerable from the approvals inbox, whichever providers are configured.
Output
The agent's result, together with the run's telemetry. The result's shape is defined by the agent.