Governance
The hardest part of letting AI agents do real work isn't the AI. It's trust. Here's how we're earning it.
Every agent has a budget.
The system seeds an org-chart of agents — CEO, Verifier, Research, Marketing, Finance/CFO, Support, Growth Analyst — each with a monthly USD spend cap. Cross the cap and the agent auto-pauses; the cycle orchestrator files an approval asking you to bump or wait.
Every action is appended to an immutable log.
The audit_events table is append-only at the database layer (UPDATE and DELETE both throw). Every chat turn, every task completion, every approval, every spend rolls in as a row. Download the whole log as JSON anytime.
Risky actions stop and ask.
Send external email, post to your social channel, bump a budget, declare a pivot — all gated by typed approvals. Approve in chat, in Slack, or via API. Every resolution is paired with an audit event.
Failed work is auto-refunded.
The Verifier agent independently scores task completion against the original verification strategy. If it fails, the cycle runner refunds every usage_event tied to that task and rolls the credit back. No emails, no support tickets.
Your data is yours.
One click in Settings produces a ZIP of every document (as Markdown), every chat message, every audit row, every artifact metadata blob, and the full org chart. Take it and go.
Bring your own agent.
Swap any kigya agent for the Claude Agent SDK, OpenAI Codex, Cursor, or any OpenRouter-compatible model. Scale and Enterprise tiers ship a Paperclip-compat heartbeat so self-hosted Paperclip orgs can hire a kigya agent over their own protocol.
Reading the open-source plan is the fastest way to understand the full system. Most of the design lives in docs/01-LOGIC.md and docs/05-COSTS.md.