Our first agent department. One objective, agents that own it, and a judge that sends work back until it is right. It runs when we start it, and this page updates itself after every run.
Every partner who sells with us always has current, true, on-voice material to sell with, and it stays current as the company changes.
Nobody writes that material by hand anymore. When the company changes, a price, who we serve, what we offer, the department notices, rebuilds what each partner needs, and has it judged before anyone sees it. Everything it makes is a draft. A person reviews it, and the partner sends it in their own name.
How it is built
Yellow is the lead, green the agents that make the work, blue the judge. Dashed lines are the loops: a kit goes back until it passes, and Dreaming tidies memory between runs.
LeadReads what changed since the last run, decides what each partner needs, briefs the specialists, checks their work and owns the result. A large model, working as coordinator.
MakerBuilds or updates one partner's kit: the page they send, the play they run, the first-meeting talk track, follow-up drafts and a researched target list. Several run in parallel.
KeeperWhen a fact changes anywhere, it changes everywhere. The Keeper finds every place a price, a term or a promise appears and makes them agree.
JudgeA separate grader with its own context scores every kit against a written rubric: our voice, the Live Solve offer, what each kind of partner needs. A kit that fails goes back with the reasons. Fail means fix and resubmit, not escalate.
MemoryOne store for the department, one for each partner: the facts every kit must agree with, what each kit says, what changed. Small files, so they can be consolidated.
DreamingAfter a run, consolidates the department's memory: merges duplicates, keeps only sourced facts, folds old notes into a short history. It ran for the first time after the first run.
What it runs on
We name the parts honestly, including what is new and still changing.
Claude Managed Agentsbeta: Anthropic runs the agent loop and the sandbox each agent works in.
Outcomesbeta: the judge. A grader scores the work against a rubric and the agent revises until it passes.
Multiagent orchestrationbeta: the lead delegates to specialists, each in its own thread, in parallel.
Memory stores beta: what the department remembers between runs, versioned.
Dreaming research preview: memory consolidation after each run.
The shape is not ours. It is two patterns from Anthropic's Building effective agents: orchestrator–workers (the Lead and its specialists) and evaluator–optimizer (the Judge and the revise loop). We applied them to one objective in our own company.
The rules it works under
Drafts only. Nothing goes to a partner or anyone outside without a person.
Every fact comes from our own documents. A missing fact is marked open, never guessed.
Every run is logged with what it cost, here and on our own Bridge.
Partner names, prompts and rubrics stay inside the company. This page shows the numbers.
The record
Being built. The first run is under way; its numbers appear here once the judge has graded every kit.