One lesson, from one agent to the whole team
Fernhill is an online plant shop with four agents and the four people who answer for them. On a Monday one of the agents finds out why an order arrived wilted. Scroll to follow what happens to that one piece of knowledge, and where a person steps in.
Fernhill is made up. The names, memories and numbers are sample data.
-
1. Each agent connects once
An agent connects through MCP, the standard way agents use tools. It gets a part of the circles that is its own, and a person who answers for it. An agent can never see more than that person can.
- ops-agent · Orders, shipping and refundsDaniel, Operations lead
- support-agent · Customer email and chatMaya, Support lead
- sales-agent · Wholesale quotesPriya, Sales lead
- code-agent · The shop's website and checkoutTom, Engineering lead
-
2. A session starts, and the agent already knows things
Before ops-agent reads its first instruction, it is handed its slice: what it learned before, the team's lessons and the decisions that apply. It did not have to ask, and no model was called to put the slice together.
- Refunds over $500 need a person's approvalDecision
- Quote delivery dates from the warehouse calendar, not the storefront bannerTeam lesson
- The Friday courier pickup closes at 11:00Its own memory
- Warehouse stock counts update at 06:00 and 18:00Its own memory
-
3. It works, and takes notes
Today's task is to find out why order 48213 arrived wilted. As it works, the agent keeps notes on the outer circle. They belong to this task and are gone two weeks after it closes.
- Order 48213 was placed Thursday 19:40 and picked up Friday.
- It sat in the courier depot until Monday. The courier does not move live plants over the weekend.
- Refunded under the 14 day rule.
-
4. It keeps what is worth keeping
One of those notes will matter again, so the agent remembers it. From now on ops-agent knows this in every session. No other agent does yet.
- The courier does not move live plants over the weekendOnly ops-agent
-
5. Every new memory is checked
The moment a memory is written, it is compared with what the team already holds. There are two ways this goes.
See the contradictions tableNothing disagreesThe courier does not move live plants over the weekend
Saved. This is what happens almost every time.
Something disagreesPlants can be refunded within 30 days, written by support-agent, against the decision "Live plants can be refunded for 14 days after delivery".
A private memory loses to a decision in force. The memory is archived and the agent is told why. When no rule fits, the pair waits for a person and no agent acts on either side.
-
6. It proposes the lesson to the team
The agent thinks everyone should know this. It can propose, and that is all it can do. The draft stops at the dark ring. The agent gets a receipt and cannot read the draft back.
- The courier does not move live plants over the weekend. Orders placed after Thursday noon ship on Monday.Waiting for a person
-
7. A person approves it
Daniel owns this part of the team's memory, so the draft is in Daniel's queue. Approving takes one step. Rejecting takes a reason, and the agent gets it.
See the review queue -
8. The next agent starts with it
The next morning support-agent opens a session. It has never worked on shipping, and the lesson is already in its slice.
- The task
- A customer asks when a Thursday evening order arrives
- What support-agent answers
- It ships Monday and arrives Tuesday.
- Without the lesson
- It would have promised Saturday.
-
9. You can check what it was worth
Every task is recorded: whether it really passed, what it cost and which slice the agent was given. Run the same tasks cold and with memory, and you get the two side by side. Passing is decided by tests the agent never sees, and cost is measured from outside the agent.
Sample numbers from a made-up team. Yours come from your own tasks.
See the records you get