Make authority explicit and turn judgment into systems
The pattern
Oddlynew works best when authority is explicit. People and agents can gather evidence, propose interpretations, exercise judgment, and make decisions when they are authorized and capable of doing so. What matters is who has the relevant context, who represents or bears the consequences, and which safeguards the decision requires. Speed, substrate, or confidence alone does not confer authority.
The useful pattern is a loop rather than a fixed assignment of human and agent roles. An authorized participant assembles evidence and proposes a direction. The accountable decision owner decides, delegates, or requests more work. The decision and its reasoning return to the workspace, and later outcomes correct the shared model. In one workflow the decision owner may be a person; in another it may be an authorized agent or a group. The boundary follows responsibility and consequences, not substrate. When the owner is a person, the proposal reaches them as a question whose answer is kept where the next agent finds it, as Ask a person, and keep the answer describes.
Why the decisions belong in the workspace
A decision that lives only in a chat window evaporates. A decision recorded as a document, carrying what was decided, on what evidence, and against which alternatives, does three things over time.
It makes the next decision faster. An agent can surface how you decided similar questions before, and what happened afterwards.
It makes judgment auditable. Wrong calls stop being embarrassing and start becoming evidence for the next decision.
It becomes a system. When a kind of decision has been made well several times, the pattern becomes visible enough to codify: the structure of a document type, a checklist, or a standing instruction that an authorized agent or person can follow for routine cases while consequential exceptions remain with the appropriate decision owner.
That last step is the compounding move. Good judgment, exercised, recorded, and tested against outcomes, gradually becomes replicable structure. Participants spend less attention re-solving settled cases and move their judgment toward what is genuinely new.
What this looks like concretely
An operator connects their ad platforms, CRM, and analytics, and asks their agent for a regular pulse: everything fetched, compared, summarized. They read it, decide, and have the agent record the decision and the reasoning as a document.
Over weeks, the decision documents show a pattern, including which signals they trust and which recommendations they routinely reject and why. The agent drafts a standing evaluation guide from them; the person corrects it. Routine calls now follow the guide, and unusual ones still come to the person.
When results diverge from expectation, the trail exists: what was believed, what was decided, what happened. The correction goes into the guide. The system gets better because the judgment did.
The boundary
None of this prescribes one universal division of labor. Some work is delegated entirely, some remains closely reviewed, and the boundary changes with capability, trust, authorization, and impact. Oddlynew supports co-evolution: people and agents think better together when evidence, decisions, corrections, and responsibility remain visible.