SoftAgent

The control runtime for AI agents that act with money.

Plan. Approve. Execute. Verify. Recover. SoftAgent는 결제가 아니라 controlled agent execution입니다. LLM이 게이트를 판단하지 않고, 승인된 계획만 실행하며, 검증 없는 capture는 완료가 아닙니다.

Demo (5분)

  1. Workbench에서 Approval Demo — paid tool이 `WAITING_FOR_APPROVAL`에서 멈춥니다.
  2. 같은 화면에서 Approve and resume 또는 Approvals inbox.
  3. 완료되면 Runs에서 DecisionTrace · public refs (pi_/rcpt_)를 확인.
  4. Verification Failure Demo — capture 없이 auth release / FAILED_SAFE.
  5. Integrations — Soft*는 optional adapter, `/ready` blockers.

실사용

  1. Agents에 실행 주체를 등록하고, Policies에서 delegation + approval threshold를 겁니다.
  2. Host는 `POST /v1/intents`에 delegation_id만 넘깁니다. SoftAgent가 store에서 hydrate합니다. 키는 Console 서버 프록시 뒤에 있습니다.
  3. 한도 초과 run은 Approvals에서 structured scope로 승인합니다. 승인 후 같은 runId로 재개됩니다.
  4. crash/timeout은 실패가 아닙니다. Run 상세에서 Reconcile 또는 Compensate — unknown capture는 환불하지 않습니다.
  5. 연동은 Public API · Public ID · Signed Webhook · Idempotency-Key만. DB 공유 금지.

Why this wins

RunsApprovalsPoliciesIntegrationsWorkbench demoAgents