Working notes from the people who build the agents
September 9, 2026 · Reqmint
An agent is only as current as its copy of the ATS. This walkthrough builds the sync layer underneath one: an event ledger, webhook signature checks, deduplication, a nightly catch-up pull, and a staleness number you can put on a dashboard.
Read moreSeptember 9, 2026 · Reqmint
Logging every automated screening decision is the easy half. This walkthrough builds the other half: a retention table with a clock per record class, partitioned decision logs, legal holds that beat the purge job, and a purge that leaves proof it ran.
Read moreSeptember 9, 2026 · Reqmint
A screening agent tested once, by hand, drifts the moment a model version or a prompt changes. This walkthrough builds the harness: a frozen golden set, pinned model and rubric versions, an agreement and movement diff, and rationale checks that fail the run.
Read moreSeptember 9, 2026 · Reqmint
Approved hours, the pay run and the invoice run drift apart every cycle, and the difference is margin. This walkthrough builds the reconciliation: three landing tables, a rate card with effective dates, five exception rules in SQL, and a Tuesday exceptions list a human works.
Read moreSeptember 9, 2026 · Reqmint
A finishing-soon list tells you who to call, not what about. This walkthrough builds the matching step: a skill alias table, hard filters before scoring, a four-component score with tunable weights, stored rationale per pair, and the narrow job a model should actually do.
Read more