Bind

@Jaydearcadian·Jaydearcadian·15 followers·added 2026-10-24

Bind puts an agent team inside a real business workflow. A commercial-kitchen vertical slice runs procurement, fulfilment and menu agents whose every tool call is classified as read, external_request or commercial_commitment by src/policy/effect-classifier.ts, and src/policy/engine.ts decides allow, deny or require-approval from workspace spend limits and counterparty allowlists before any model output can commit funds. Workflows move through explicit states (draft → request_sent → quote_received → awaiting_approval → authorized → submitted → confirmed → reconciled) in a PostgreSQL-backed engine (src/workflows/engine.ts, src/workflows/recovery.ts) with idempotency keys and crash recovery, and src/accounting/three-way-match.ts matches purchase order, receipt and invoice in integer cents. The optional deepagents LLM loop only picks a quote and writes rationale. On chain, contracts/src/BindSettlement.sol enforces a rolling daily limit and vendor whitelist, records order ids and match hashes, and exposes depositEscrow, releaseEscrow, refundEscrow and settleOrder, deployed on Arc testnet at 0x16E1Bd42F7C0FfD28D10919C87CaD22d05f2aA29.