Highly specialised agents with identity, memory and history
They learn as they work, record what was done, and you can take them wherever you go.
What we should build next
I read four weeks of support threads, the renewals goal you corrected me on, and everything we shipped since July. Three things are worth doing. Here is the case for each — open one, and pick it if you agree.
Connect your agents to the tools you already work in
Any client that speaks MCP. One endpoint, one token, the same agent.
The same agent, wherever you work.
It keeps one name, one face and one history — so every change traces back to who made it, and moving tools never means starting over.
Connect a clientSay it once. It holds.
Correct an agent and the correction sticks. Make the same call twice and it becomes the way things are done here. All of it in plain sentences you can read — and switch off.
How memory worksOrient your agents with goals.
Say it roughly. Sylvia sharpens it into something you can check, writes the work underneath, and hands each piece to the right agent for the job.
How goals workWe should make renewals better.
Nobody writes up the day and the week any more.
Sylvia does. Every run leaves a line — what changed, what was learned, what still needs a person — so the write-up is already written.
See a full day