A governed content-editor that turns upstream intelligence into grounded, attributable drafts — proposed to humans, shipped by no one but them.
Every other agent produces an artifact you can quietly revise. The content-editor produces words that carry the company's voice — and a model that both writes and ships is a brand incident with no seatbelt.
A drafting model with a publish key. No provenance on what it grounded on, no human gate, no signed record of who authored the claim. Fast — and unaccountable.
Barbara owns messaging. Agents propose; humans publish. Every draft is grounded, attributed, policy-checked at the enforcement plane, and written into an audited, tenant-isolated store.
The content-editor is not a lone assistant. It is one governed tenant inside a live fleet mapped like-for-like to the real ACME organization.
The content-editor takes finished analysis from upstream and turns it into a draft — then hands that draft down to compliance. It never originates raw intelligence, and it never closes the loop alone.
Nothing the content-editor writes reaches the store without passing the enforcement plane. Policy is deny-by-default; on allow, the call is forwarded verbatim to this agent's own isolated data tenant.
This is not a mock. The content-editor runs on a schedule, homes to a free tier, and every job lands in the usage table behind the Grafana ROI dashboard.
One scheduled job, start to finish. The editor does not free-associate — it consumes a domain brief and produces a chain-of-thought and a memory write that both carry their sources.
The editor homes low to keep routine drafting near-free, but the same policy that keeps it cheap also lets it climb when the material demands it — and degrade safely when a model fails.
The write is the moment governance bites. The enforcement plane validates the caller, the data plane validates the bearer, and the store refuses anything unattributed.
source_type — what it grounded on. authored_by — which agent produced it. A tenant bearer hashed as sha512 against _soul_api_keys. Without these, ContentInspection (enforce) blocks the write.
The agent's own row-level-security tenant in acme_soul. _memories and _soul_api_keys run FORCE RLS, so one editor can never read or overwrite another agent's drafts.
This is the whole design in five words. The content-editor's output is a proposal that a person reviews and releases — the machine never crosses that line on its own.
The ROI board answers the self-host-vs-API question with observed per-call cost. For a high-volume drafting agent, homing to a free tier is not thrift — it is the difference between viable and not.
| Model | Observed cost / call | Relative |
|---|---|---|
| opus-4.8 (paid) | $0.19 | |
| sonnet-4.5 (paid) | $0.03 | |
| deepseek-v4-pro (Ollama) | $0.002 | |
| gpt-oss:120b (Ollama) | $0.0015 | |
| Llama-3.3-70B (HF) | $0.0012 | |
| gemma-4-31b:free | $0 |
The content-editor makes drafting fast and near-free without ever handing a model the last word. Every claim carries its source, every call is signed, and Barbara still owns the moment it goes live.