Agents · signals · data · one operator
A social swarm of AI agents. One human at the gate.
Thousands of small agents read the signals, find each other, argue and draft. Nothing leaves the swarm until a person says yes.
Fifteen minutes inside the swarm.
One call. No slides. A live look at the agents working, the gate where the operator decides, and what happens to the replies. Then you decide whether you want in.
Invitations open soon. Watch this frequency.
- SwarmMany agents, not one chatbot.
Scouts find the signals, writers draft, critics push back. They work as a group.
- GateA person approves every send.
The agents propose. One human reads, edits and decides. That step is never automated.
- LoopReplies feed the next cycle.
What comes back goes into the next run, so the swarm learns who answers and why.
Data · under the swarm
Swarms run on data. The data gets a gate too.
Every agent reads and writes structured data. Before anything reaches the operator, that data has been checked, indexed and scored, and the results get measured.
-
01 · Schema
Checked before it's written
Every record is validated against a schema before an agent may write it. Bad rows are refused, not quietly repaired. History is appended, never overwritten.
Running -
02 · Index
Read-only, whitelisted
Agents index private notes without changing a byte. A field whitelist decides what enters the index, so contact details never do. Tested on synthetic data only.
Running -
03 · Match
People to briefs
Language, region, seniority and skills turn into clauses, and the bench is scored against every open brief. Rules today. Embeddings and clustering are next.
Running -
04 · Measure
Did the swarm get it right?
Each approval, edit and rejection by the operator becomes a label. Reply rates, shortlist precision at k and drift show whether the swarm is getting better or just getting faster.
Blueprint
Field manual · swarm types
Twelve swarms. One pattern.
Each one is the same loop with different agents: scouts find, writers draft, critics check, one human decides. The first ones were built by one operator in seven days. Pick one for your fifteen minutes.
- Running in daily use
- Prototype built and tested, not in daily use
- Blueprint designed, not built
-
Running
Recruiting desk
Scouts read the brief and the bench. Writers draft the shortlist, client-ready CVs and outreach. Critics check every draft against the house rules: length, one ask, an opener backed by evidence, consent for the channel.
Gate · the operator approves the exact text
-
Running
Inbox triage
Reads chat and social inboxes inside your own browser, finds who is owed a reply and how long they've waited, and proposes the tracker update.
Gate · read-only, it sends nothing
-
Running
Chief of staff
The morning brief: deadlines, what's due or overdue, the next two working days. Logs what happened in one line and keeps the history. Scheduled runs report back.
Gate · writes only when you ask
-
Running
Knowledge
Indexes a private notes vault without changing a file, matches people to open briefs on language, region and seniority, and lists the data gaps. Contact details never enter the index.
Gate · read-only, the vault stays the source of truth
-
Running
Builders
Agents work in parallel cloud sessions, leave each other handoff notes and open pull requests with tests. CI and a reviewing agent check the work before anyone looks.
Gate · you merge
-
Running
Studio
Made this page: the film, its score synthesised from scratch, the site and its checks for privacy, page weight and banned words. Several cuts from one source.
Gate · you pick the cut, nothing publishes without you
-
Prototype
Business development
Maps the decision-makers at one account, drafts the note, the fallback message and the follow-up, then sends, schedules and logs in one pass.
Gate · one approval per account, on the exact texts
-
Prototype
Security watch
Scans a machine for persistence against a baseline, watches which programs talk to the network and flags upload spikes while the screen is locked. A kill switch cuts outgoing traffic and restores it exactly.
Gate · the kill switch is yours; automatic only if you opt in
-
Blueprint
Market signals
Watches hiring, funding and job-post signals across sources, scores accounts on fit and timing, and writes a weekly brief on who to call and why.
Gate · you choose who gets contacted
-
Running
Data quality
Scans a data store's structure (field names, types, counts, never the values), validates rows against schemas and lists the gaps that would break a match.
Gate · it reports; you fix the source
-
Running
Analytics
Turns the index into one page: target against actual, the pipeline by stage, coverage against demand, what's due. Rebuilt in seconds, no data leaves the machine.
Gate · read-only
-
Blueprint
Model evaluation
Scores the matcher and the drafts against the operator's own decisions, tracks precision at k and reply rates over time, and flags drift before it shows up in results.
Gate · you decide when a model changes