Software Agents Quickstart

Software-engineering agents are great at creating code, changes, and pull requests — but most teams stop them before operational actions like merging, deploying, running migrations, rotating secrets, or destructive operations. BobSentry is the runtime authorization layer for that boundary: call POST /v1/runtime/evaluate with x-api-key, get ALLOW, REQUIRE_APPROVAL, or BLOCK, then honor the decision in your own execution path. BobSentry does not execute the downstream action.

1. Get a sandbox key

export BOBSENTRY_SANDBOX_KEY=bsk_your_key_here

2. Evaluate an action

Call the runtime before your agent performs the action. The action field selects the policy; the decision is deterministic.

curl -X POST https://runtime.bobsentry.com/v1/runtime/evaluate \
  -H "x-api-key: $BOBSENTRY_SANDBOX_KEY" \
  -H "content-type: application/json" \
  -d '{
    "action": "deploy_production",
    "agent_id": "infra-agent",
    "action_environment": "PRODUCTION",
    "schema_version": "2026-06",
    "metadata": { "system": "payments" }
  }'

Action catalog

Five software-engineering actions ship in the runtime taxonomy today. Each example below is runnable with your sandbox key.

merge_pull_requestALLOW

Policy SE-PR-MERGE-001

Request

{"action":"merge_pull_request","agent_id":"agent_demo",
 "metadata":{"repo":"acme/api","branch":"feature/login"}}

Response

{"decision":"ALLOW","risk_level":"LOW",
 "policy_id":"SE-PR-MERGE-001","approval_required":false,
 "approval_status":"NOT_REQUIRED","signature":"ed25519:…"}

Why: Merging a reviewed PR into a non-protected branch is low blast radius, so it runs without a human in the loop — but every decision is still logged with signed evidence.

deploy_productionREQUIRE_APPROVAL

Policy SE-DEPLOY-001

Request

{"action":"deploy_production","agent_id":"infra-agent",
 "action_environment":"PRODUCTION","schema_version":"2026-06",
 "metadata":{"system":"payments"}}

Response

{"decision":"REQUIRE_APPROVAL","risk_level":"HIGH",
 "policy_id":"SE-DEPLOY-001","approval_required":true,
 "approval_status":"PENDING","signature":"ed25519:…"}

Why: A production deploy has operational impact. BobSentry returns REQUIRE_APPROVAL so the integrator must not deploy yet; a release manager authorizes in the console, then the integrator may proceed.

run_database_migrationREQUIRE_APPROVAL

Policy SE-DB-MIGRATION-001

Request

{"action":"run_database_migration","agent_id":"agent_demo",
 "metadata":{"database":"prod","migration":"20260624_add_index"}}

Response

{"decision":"REQUIRE_APPROVAL","risk_level":"HIGH",
 "policy_id":"SE-DB-MIGRATION-001","approval_required":true,
 "approval_status":"PENDING","signature":"ed25519:…"}

Why: Schema and data migrations can be irreversible, so the action requires human authorization before it is applied to a production database.

rotate_secretREQUIRE_APPROVAL

Policy SE-SECRET-ROTATE-001

Request

{"action":"rotate_secret","agent_id":"agent_demo",
 "metadata":{"secret_id":"prod/api-key","scope":"production"}}

Response

{"decision":"REQUIRE_APPROVAL","risk_level":"HIGH",
 "policy_id":"SE-SECRET-ROTATE-001","approval_required":true,
 "approval_status":"PENDING","signature":"ed25519:…"}

Why: Rotating a live credential can cause outages or lockouts, so it is gated on explicit authorization. Only the secret identifier is recorded — never the value.

delete_repositoryBLOCK

Policy SE-REPO-DELETE-001

Request

{"action":"delete_repository","agent_id":"agent_demo",
 "metadata":{"repo":"acme/legacy"}}

Response

{"decision":"BLOCK","risk_level":"CRITICAL",
 "policy_id":"SE-REPO-DELETE-001","approval_required":false,
 "approval_status":"NOT_REQUIRED","signature":"ed25519:…"}

Why: Deleting a repository is irreversible loss of source history. It is blocked outright with no automated override path, and a signed denial record is produced.

What happens on REQUIRE_APPROVAL

When a decision is REQUIRE_APPROVAL, the runtime returns approval_status: "PENDING" and opens an approval ticket. Do not execute the downstream mutation. A human authorizes in the BobSentry console; poll GET /v1/runtime/decisions/:decisionId until approval_status is terminal. Evaluation credentials cannot self-approve. ALLOW means the integrator may proceed; BLOCK means stop. See Approval Routing and the Runtime Integration Contract.

Evidence

Every decision returns a decision_id (also the evidence id) plus an Ed25519 signature and a hash-chain digest, so any action your agent takes is independently verifiable. The tenant is derived from your API key on the server. Recent decisions appear in your Sandbox evaluation history.