Case Study

They stopped doing the work. Now they approve it.

A multi agent content system that lives in Slack. It keeps the calendar full, researches, drafts, reviews its own work, and files the finished piece to Notion with a full audit trail. A person approves every step.

  • Client. Marketing and design startup, about 60 people
  • Slack
  • Notion
  • Claude Agent SDK
  • Status. Live in production

One run, start to finish

This is what a piece of work looks like now.

  1. 01

    The ask

    The old shape of the work: research the topic in a browser, write the draft, paste it into Notion in the right place, then log what was done. Several times a week, across three brands.

    Marketing leadHuman

    run the blog workflow for the spring launch, audience is ops leads

  2. 02

    It picks the playbook

    Workflows are Notion pages, not code. The team writes new ones themselves, so the system grows without waiting on a developer.

    AgentOrchestrator

    Loaded the blog generation playbook. Twelve steps, with approval points along the way.

    Confirming brand, topic, audience and publish date before I start.

  3. 03

    It reads before it writes

    One agent, and only one, is allowed to read Notion. It returns source material verbatim, without interpreting it, so nothing is invented on the way in.

    AgentFetcher

    Pulled brand strategy, the ICP, voice guidelines and competitor references.

  4. 04

    It checks it is not repeating itself

    Every brand accumulates near duplicate posts over time. The agent checks the existing library and the dedup ledger before writing a word.

    AgentOrchestrator

    This overlaps an existing post. Proposing a different angle instead.

    Want me to go ahead with the new angle?

  5. 05

    It drafts, then argues with itself

    The validator reviews the draft against the source material and returns approve or revise with specifics. That loop runs inside the agent team, so a person never sees a raw first draft.

    AgentValidator

    REVISE. Some claims are not supported by the retrieved sources, and the intro drifts from the brand voice.

    Redrafted and rechecked. Approved on the second pass.

  6. 06

    A person says yes

    This is the whole safety model. The agent pauses and asks in the thread, and the run stops there until someone answers. Gates ship on by default, and only the client decides if the agent has earned looser ones.

    AgentOrchestrator

    Draft is ready for review.

    Waiting on a person

    • Approve
    • Send feedback
    • Cancel

    Feedback sends the step back to be redone. Nothing is written anywhere until someone approves.

  7. 07

    It files the work and the receipts

    Alongside the deliverable the agent keeps a task log in Notion, a step by step record of exactly what it did. Publishing stays a human step. The agent stops at ready to publish.

    AgentOrchestrator

    Filed the deliverable, wrote the task log, marked the card ready to publish.

Under the hood

Four agents, each with one job.

Splitting the work across specialists is what makes it affordable and what makes it checkable. The cheapest model that can do a step is the one that does it, and the most capable one is spent on review.

Orchestrator

Runs the thread

Receives the Slack message, steps through the playbook, coordinates the three specialists, and manages the approval cycle.

Fetcher

Fast model

Retrieval only, and the only agent allowed to read Notion. Returns source material verbatim, without interpreting it.

Analyzer

Mid tier model

The writer. Drafts the piece, prepares the metadata, writes the art direction brief, and assembles the handoff package.

Validator

Highest capability, used sparingly

Quality control before a person is involved. Returns approve, or revise with a specific list of problems to fix.

The calendar keeps it fed

A second workflow acts as the planner. It keeps each brand's content calendar populated, runs dedup across the whole calendar rather than one post at a time, and dispatches the production workflow when a scheduled slot comes due. The team asks for a batch of content instead of a post.

How it went

A repeatable engagement, not a black box.

  1. 01

    Audit

    Under two weeks, four working sessions to map the bottleneck.

    We watched the real workflow and found the one that ate the most time.

  2. 02

    Build

    Built, deployed and tested in two weeks.

    Built against their actual Slack and Notion setup, with approval gates from day one.

  3. 03

    Live

    Running in their workspace, tuned with their feedback.

    Doing real work, refined from how the team actually uses it.

Your team has a workflow shaped like this one. The first step is finding out which one is costing you the most.

Start with the €500 Audit