An autonomous media buyer
that asks a human before touching the budget
Our own product in development: an AI agent that watches Meta ad accounts, diagnoses problems, proposes and executes changes within approved limits. Scheduler, orchestrator loop, rules engine with Claude for ambiguous cases, idempotent executor with snapshot and rollback, approval queue in Telegram, four-level trust ladder, three circles of authority, two-level budget protection and a kill switch. Six architecture documents and a code skeleton.
The idea
A media buyer’s day is a loop: look at the numbers, find what broke, decide, change, check. Most of it is rules; some of it needs judgement. The agent runs the loop, applies the rules, asks a model for the ambiguous cases, and asks a human for anything that moves money.
The guards
A trust ladder from observation only to auto-publish, earned per account. Three circles: actions the agent may do alone, actions that wait for approval in Telegram, and actions it may never do (billing, access, deletion). A soft budget cap in software plus a hard spend cap physically set on every ad set in Meta, without which publishing is blocked. Snapshot and rollback on every change, idempotent execution, a kill switch, dry-run by default.
Status
Six documents and a code skeleton (brain, config, rate-limited Meta client, rules engine, orchestrator, executor, Telegram bot), with guardrail configs for two real accounts. Clients on our management packages receive early versions of this as it matures.