Skip to content
Mark Warren
Menu

All work

Approval-first agents

Agents that draft, with a person in the loop

An operating system for a small professional-services company, built around one rule: nothing leaves the building without a person's approval.

How it fits together

  1. Inputs

    • Inbound mail

      A scheduled poll reads new threads, with a hard cap on how many one run can touch.

      • Gmail
      • n8n

    • Time entry by text

      Plain-language messages from a phone go to one shared bot that serves every agent.

      • Telegram

    • Web leads

      Form posts reach a webhook that checks a header secret and a honeypot before anything runs.

  2. Engine

    Workflow engine

    n8n workflows with Postgres state on Supabase. Every business fact comes from one version-controlled source.

    • n8n
    • Supabase
    • PostgreSQL
    1. Classify and parse

      A smaller Claude model returns structured fields through forced tool calls. If the classifier fails, mail passes through instead of being dropped.

      • Claude
    2. Match without guessing

      Client names are matched by rules. An ambiguous match asks which client was meant.

    3. Draft

      A stronger Claude model drafts replies, invoices, and posts. No workflow has a send step.

      • Claude

  3. Outputs

    • Approve in Telegram

      Each draft waits until a person approves, edits, or rejects it.

      • Telegram

    • Draft invoices

      Approved time becomes a draft invoice in QuickBooks Online, never a sent one.

      • QuickBooks Online

    • Lead records

      Qualified leads are recorded in HubSpot.

      • HubSpot

Guardrails

  • Monthly spend limit and a daily spend watchdog
  • Hard cap on items per run
  • Tests on every push and a daily health check
  • A second model reviews every plan before it is builtOpenAI

The problem

A small services company spends its week on intake, invoicing, call follow-up, and content. Nearly all of it goes to clients or the public, and a mistake is visible. Automating that work is easy. Automating it safely is the job.

What I built

  • Agents that triage inbound mail and draft replies, invoices, and posts, with no send step that a person does not trigger.
  • Time entry by plain-language text message that becomes a draft invoice, and asks which client was meant instead of guessing when a name is ambiguous.
  • Public-facing agents with authenticated entry, hard limits on how many items one run can touch, and a classifier that fails open so a failed model cannot swallow real mail.
  • Spend controls for unattended runs: monthly limits, a daily watchdog, and a hard limit on each run.
  • Automated tests on every push to every site, and a daily health check that exercises the whole submission path.

How I built it

The workflows run on n8n with Postgres on Supabase, Claude models for parsing and drafting, and Telegram for approvals. A smaller model handles classification and parsing through forced tool calls, and a stronger model drafts. A second model reviews each plan before I build it, and every business fact lives in one written source that the agents and the sites read from.

What it shows

I treat approval, spend, and failure behavior as design requirements from the first day, not as cleanup after launch.