recommendAgentPolicy

Derives the tightest policy that would still pass every statement this agent credential has legitimately run, using its recorded audit history over a lookback window (default 30 days). The candidate's table allowlist is the union of relations actually referenced, its statement-kind allow set is the observed set, it downgrades to read-only when no writes were seen, and its max_rows ceiling comes from an observed high-percentile row count. The candidate is proven safe by replaying it through the data plane's own policy engine against the same history: a good recommendation has replay.summary.newly_blocked == 0. This endpoint is advisory only. It reads the audit log, never connects to the upstream database, and never creates, updates, or mutates any policy or credential; the operator loads the candidate into the editor and saves it themselves.

Usage

const result = await api.policies.recommendAgentPolicy({
    pathParams: { project_id: "prj_xxx", agent_id: "agt_xxx" },
    body: {
      ,
    },
  });

Parameters

Response

Promise<PolicyRecommendation>: the recommended candidate policy and its replay proof.

Example

import { PgBeamClient } from "pgbeam";

const client = new PgBeamClient({
  token: "pbk_...",
  baseUrl: "https://api.pgbeam.com",
});

const result = await client.api.policies.recommendAgentPolicy({
  pathParams: { project_id: "prj_xxx", agent_id: "agt_xxx" },
  body: {
      ,
    },
});

Errors