Info@blueism-studio.com — Online worldwide · HQ in Tetouan, Morocco
The brief
{{ briefBody }}
Approach
{{ a.n }}
{{ a.body }}
Plates
Drag sideways to turn the page
01 — Selection
Before building anything we sorted the week on two axes: how often it recurs, and how much judgement it takes. Only one quadrant is worth an agent.
02 — Architecture
One n8n workflow per agent. Every one of them runs the same five-stage spine, so a new agent is a configuration rather than a project.
01 Trigger
Wake up
Cron — Mon 06:00. Gmail watcher. Inbound webhook.
02 Gather
Assemble context
Airtable records, Postgres history, Drive and mail threads.
03 Reason
Form an opinion
Claude, given the rubric and prior decisions. Returns a reason and a score.
04 Gate
Act or ask
Confidence at or above threshold acts. Anything below escalates.
05 Act
Do the thing
Write to Airtable, draft in Gmail, call an internal API.
Fallback
A Slack thread to a named human, with the reasoning attached.
Run ledger
Every run writes input, reasoning, decision, confidence and outcome to Postgres. Nothing an agent did is unexplainable after the fact.
03 — The roster
Each one is narrow on purpose: a single decision, a single rubric, a single person it can interrupt.
Agent 01
The Reconciler
Reads the week’s movements against what was expected and reports only the differences that need an opinion.
Airtable → Postgres → Slack
Agent 02
The Triage Clerk
Sorts the inbound queue by what a message is actually about, not by what its subject line claims.
Gmail → Claude → Airtable
Agent 03
The Follow-up
Notices what has gone quiet, drafts the nudge in the right register, and leaves it for a human to send.
Postgres → Claude → Gmail draft
Agent 04
The Registrar
Keeps the record straight: fills the gaps it can prove, flags the contradictions, and never invents a value.
Webhook → Claude → Postgres
04 — Method
An agent allowed to decide needs three things a script does not: a rubric, a memory, and a way of saying it is not sure.
The rubric
The policy the team already had in their heads, written out as ordered rules. Most of the project was this conversation.
The memory
Prior decisions are retrieved and passed in, so this week is judged the way last week was.
The gate
Low confidence is a valid answer. It routes to a person instead of guessing.
Rubric — extract
DECIDE: route | hold | escalate05 — Outcome
Monday begins finished
The recurring run completes before the first person logs in. What is waiting is a short list of held items, not a full queue.
Exceptions became the job
People spend their attention on the items the agents could not settle — which is the part that needed a person all along.
The record stopped drifting
Because every decision is logged with its reasoning, the history is now answerable. Disagreements are settled by reading, not remembering.
Grading
Drag across the image
Delivered
{{ s.label }}