In short
Lobster Herding is PostScarcity AI's service for running a fleet of AI agents in production: one lane per agent, one board for every task, schedules, human sign-off on anything that leaves the building, budgets, and a record of every run. The client owns the files, on any harness and any model.
01 · Why a lobster
The mascot of the agent age.
The lizard people already had the internet. The agents needed their own animal, and they picked a lobster.
The open-source agent that broke out in 2026 shipped with a lobster mascot. It renamed itself twice in one week and called each rename a molt. Today it is OpenClaw, with more than 390,000 stars on GitHub. Its agents even got a social network of their own, where only agents post and humans watch.
So the lobster became the face of agents that do things, the mascot for the run-up to the singularity. Lobsters grow by molting. Good agents do too: swap the model, keep the agent.
02 · The problem
One agent is a project. Ten is an operation.
The first agent is easy to watch. By the fifth, two of them do the same job, one failed quietly on Tuesday, and nobody can say what shipped or what it cost.
A fleet needs herding: lanes, a board, a clock, a gate and a budget. That is the service.
03 · How we herd
Lanes, a board, a clock, a gate, a budget.
Every agent has a lane.
One job, one set of keys, one owner. No agent steps on another's work or reads another's data.
One board for humans and agents.
Every task is a card. Every card ends with proof: the link, the file, the screenshot.
Recurring work runs on a clock.
Reports, posts, digests and checks run on schedule. Silent when clean, one message when not.
A human signs anything that leaves the building.
Outbound email, spend and anything public wait for an OK. Agents draft at full speed. Nothing embarrassing ships.
Agents ship under your name, by your rules.
Commits, posts and sends land under the right identity, so the audit trail stays clean.
Every agent has a budget.
Caps on every paid key and the cheapest model that does the job. Spend tracks value.
Every run is recorded.
Logs, transcripts and run history per card. A stalled job is noticed and restarted.
Every mistake happens once.
A mistake becomes a rule, a skill patch or a validator. The herd gets better every week.
Any harness. Any model. Your files.
The instructions are the code, and they are yours. Each better model is a free upgrade.
04 · The board
What a herd looks like on a Tuesday.
Finance
Monday cash report
Mon 7:00
Content
Thursday newsletter
Thu 6:00
Inbox
Triage 38 new threads
Sales
Score overnight leads
Research
Brief before the 2pm call
Inbox
14 replies drafted
Waiting on your OK
Ads
Raise spend on campaign B
Over budget cap
Content
Article published
Link + screenshot
Ops
CRM updated, 22 records
Diff attached
Support
Refund queue cleared
9 of 9 logged
05 · What is included
Built in month one. Yours by day 90.
- Instruction files for every agent, written with you and reviewed line by line
- Connectors into your systems, scoped by trust boundaries
- The board, with your team and your agents on it
- Schedules for everything that repeats
- Approval gates on outbound, spend and anything public
- Budgets per key and per agent
- Validators for anything that ships
- Documentation the agents keep current
- After day 90, if you want it: monitoring, new agents, and every move to a better model absorbed
06 · For portfolios
One playbook. Every portco.
Write the herd once, roll it across the portfolio, and report fleet-wide to the operating partner: what ran, what shipped, what it cost.
From the work
We herd our own first.
PostScarcity AI runs on a fleet of agents: content, research, sales and every client lane, each with its own files and keys, all on one board. We have moved that working fleet from one harness to another without rewriting its instructions.
FAQ