Proof of work · orchestration & decision rights

Agent & Automation Role Map

I do not treat every scheduled process as an “agent.” Known paths stay deterministic. Agentic tasks are reserved for bounded work where interpretation or synthesis adds value. Data workers supply context, outputs consolidate attention, and consequential decisions return to a human gate.

Output Agentic Task Research Agent Data / Tool Worker Deterministic Workflow Human Gate

Outputs

Follow Upone prioritized operating brief

Opportunity Briefresearch synthesis

Tech Advisordecision brief

Headhunterrole / fit brief

Marketingresearch output

Agentic Judgment

Tech Advisorbounded recommendation

Headhunterrole matching + synthesis

Opportunity Researcherevidence-led exploration

Founder Strategyscenario exploration

Data / Tool Workers

LO2.0cohort enrichment

Homebotproperty / homeowner enrichment

Remineownership context

CTIClien / title context

Deterministic

DFUdaily roster assembly

LO2 pre-7amscheduled enrichment trigger

Workflow contractsAirtable + Cloud Run paths

Human Authority

Ivanpriority · commitments · spend · merge / release

Example A deterministic when judgment is unnecessary

Enrichment Workers

LO2.0 and vendor tools refresh the context needed for the day’s operating roster.

Data / tools

→

Workflow Assembles

DFU applies known rules to due dates, birthdays, Next Touch, and available context.

Deterministic

→

Human-Ready Output

One Follow Up brief answers “who needs attention today?” without inventing a decision-maker.

Output

Example B agentic only where interpretation adds value

Retrieve Context

Bring forward current repo state, decisions, constraints, and the question to solve.

Context

→

Bounded Synthesis

Tech Advisor evaluates the defined problem and proposes one high-leverage move rather than taking unrestricted action.

Agentic task

→

Human Decision

I accept, reject, redirect, or authorize consequential execution. The outcome is written back for later retrieval.

Human gate

Operating Roster orchestration, inputs, and gates

Output	Cadence	Orchestration	Inputs / tools	Decision boundary
Opportunity Brief	Scheduled	Research agent — Opportunity Researcher	Defined research question + external evidence	Recommends; does not commit capital or launch work
Follow Up	Daily	Deterministic assembly	LO2.0 · DFU · vendor enrichment	Prioritizes context; human owns outreach and commitments
Tech Advisor	Scheduled / on demand	Bounded agentic task	GitHub state · decision history · current constraints	Proposes; human authorizes consequential changes
Headhunter	Scheduled	Bounded agentic task	Role requirements · career evidence · fit criteria	Shortlists and drafts; human decides where to apply
Marketing	As needed	Human-directed research workflow	Market research · buyer evidence · project artifacts	Research informs tests; human chooses message and spend

Design Rules the boundaries are part of the architecture

Work type	Default pattern	Escalate to human when
Known, repeatable path	Deterministic workflow with explicit inputs and outputs	Required data is missing, validation fails, or the next action changes a commitment
Research / synthesis	Bounded agentic task with a defined objective and evidence requirement	The recommendation affects spend, customer commitments, risk, or production systems
External action	Tool call or workflow after the decision is authorized	Scope is ambiguous, permissions change, or the requested action exceeds delegated authority
Learning / state	Write verified outcomes and preferences back to the shared knowledge layer	New learning would overwrite source-of-truth records or alter protected governance state

Why this matters. The architecture is intentionally mixed: agentic where judgment helps, deterministic where rules are known, and human-gated where consequences matter. That separation is a design choice, not a naming convention.

Architecture · Experience hub