How agents orchestrate work

Hakunai agents are task-oriented: they run bounded jobs inside your organisation’s permissions, not open-ended chat.

Runtime path (as implemented)#

1. Grants — Sensitive tools require an open grant with an allow-list of tool names. Empty allow-lists deny everything.
2. Policy & decide — Runtime code in lib/runtime/ evaluates proposals before side effects execute.
3. Action Centre — Humans approve or reject queued steps in the portal (/portal/action-centre) before grants are consumed.
4. Ledger / OutcomeRun — Decisions and outcomes are recorded for audit-friendly review (see docs/approvals-bridge.md in the repo).

Customer bots and internal automation use the same path — there is no bypass for staff principals (operators: see docs/BOT-RUNTIME.md in the repository).

Marketplace → portal#

• Browse agents on /marketplace or inside the portal at /portal/marketplace.
• Interest and assignment flows are live for authenticated organisations; paid checkout requires the agent-marketplace-checkout foundation flag and does not imply a live card processor.
• Portal views filter agents by suite and portal scope helpers in lib/portal-marketplace-scope.ts.

What is not live#

• Fully automated deploy-to-production for customer tenants.
• On-chain governance, $HAKU token payments, or DAO voting — future roadmap only; commercial quotes use FIAT.