Sitewarden

Autonomous AI agents that write to production systems need structured approval workflows — not post-hoc logging. The challenge is building a state graph where a human can interrupt, review, and approve before any consequential action proceeds.

The state graph runs: plan (decompose goal) → analyse (gather facts via read tools) → self-check (verify plan completeness) → write-report (draft output) → await-approval (human interrupt: approve or revise) → revise (iterate on human feedback). The risk-tiered gateway classifies each tool call: read-only tools proceed automatically; write tools (file system, API mutations) require approval before execution. Postgres checkpoints mean the graph can resume after a process restart without re-running completed steps.

1. 01

   Human interrupt approval is a first-class graph node, not an afterthought — the graph cannot proceed past await-approval without explicit human input.

2. 02

   Postgres checkpoints make the approval workflow durable: a server restart does not lose pending approvals or completed steps.

3. 03

   Risk-tiered gateway keeps read tools fast (no approval latency) while requiring explicit approval only for write operations.

4. 04

   pgvector hybrid search combines dense vector similarity with BM25 keyword matching, then attaches source citations to every retrieved chunk.

5. 05

   Retrieval evals measure recall at k; trajectory evals verify that the agent chose the right tool sequence for each test scenario.