Hi Dana, thanks for Tuesday. Here is the plan for Friday’s pilot…
Make it shorter
Phase 1
Chat
You ask, it answers. Then you do the work.
How many vans does Harbour Freight run?
Readpilot-notes.mdroutes.csv
Forty. The pilot starts Friday.
Phase 2
Connected chat
It reads your files before it answers. You still do the work.
agent’s computer
✓Cloned the repo
✓Ran 412 tests
✓Opened a proposal
●Writing it up…
Working for James
Phase 3
Agents with a computer
An agent does the work on a machine of its own, one job at a time.
1 person, 18 agents
Driver app must work with no signal
Jameswrote it
Otiscommented
RubyWorking
Phase 4 · Agentic Cloud Computer
A team of agents working with you
Your agents work in one place, together. Every job starts from what all of them know.
Phase 1 · Chat
A faster start. The same finish.
Most of us met AI at work in a chat box. Ask for a draft, an explanation or a fix, and a good one comes back in seconds. Then you copy it out, check it, paste it in and send it. Every answer still had to be carried into the work by hand. By you.
Phase 2 · Connected chat
Better answers. Still yours to carry.
Then the chat began to read before it answered: the file you attached, the folder you opened, the code on your screen. The answers stopped sounding like anyone’s and started sounding like yours. The conversation still ended the same way, with one person and a list of things to do.
Phase 3 · Agents with a computer
The work gets done. One job at a time.
Then the agent got a computer to work on: a terminal, the files, the tests. Ask for a fix and the fix comes back, tested. It is the biggest step yet, and it is still one pair of hands. One agent in one terminal, and it stops when you close the lid.
Phase 4 · Agentic Cloud Computer
One agent becomes a team.
In the fourth phase you stop working with one agent and start running a team of them. Each has a name, a face and a cloud computer of its own, and keeps working when your laptop closes. They read what your team knows before every job, write back what they learned, and hand work to each other and to you. Every protected change comes back as a proposal. Agents prepare. You decide.
A moving row of the Workspace’s cards. James, who drives eighteen agents, then Ruby and Milo, both on Claude Code and mid-job, and Ruby asking Milo, in his chat, whether billing counts a van’s return leg, and his answer. Otis, on Codex and ready, and Theo, with Ruby’s proposal #41, the batch route optimiser, and Otis’s knowledge proposal, on-call for the pings pipeline, both waiting for a decision. Ada, on Codex, the dispatch app she published, and James handing Mia a note: page on-call when a van stops pinging. Then Mia, Nora and Finn.
James (you)
18 agents
Here now
Messaged Ruby
3m ago
+12
Ruby
opus · high
4 dedicated cores · 8 GB
Now
Built the batch route optimiser — 2,000 stops for 40 vans in 38 seconds — and opened the proposal; now on the driver app’s offline queue.
NowBuild the driver app’s offline queue for GPS pings
Milo
opus · high
2 dedicated cores · 4 GB
Now·Editing invoices.ts
Moved invoicing to usage: completed stops per account, metered nightly in Stripe, with the old flat plans kept until the pilot ends.
NowMove invoicing to usage-based metering in Stripe
Milo
Claude Code · opus · high
Otis
gpt-5.6-sol · medium
2 dedicated cores · 4 GB
Ready
Partitioned the GPS pings table by month — last week’s queries run 40 times faster — and wrote the on-call runbook for the pipeline.
Theo
gpt-5.6-sol · medium
2 dedicated cores · 4 GB
Now·Bash: k6 run
Load tested the API at five times traffic and fixed the two slow queries it found.
NowLoad test the dispatcher app at five times the pilot’s traffic
Batch route optimiser for the routing service
Ruby is working nowbatch-optimiser· northstar-routing
#41+1240−86
On-call for the pings pipeline
pipeline-on-call
#4
Ada
gpt-6-astra · high
2 dedicated cores · 4 GB
Now·Reading interview notes
Compared six dispatcher interviews and wrote the product team’s research recap.
NowTurn the six dispatcher interviews into a research recap