Position A — Agents are a team you manage
- Operator: Claire Vo
- Card: Agents work when treated as a team, not a single super-tool, The unlock for AI agent productivity is management skill, not technical skill, Onboard agents the way you onboard an EA: progressive trust, named tiers
- Claim: Don't throw every task at one super-agent. Build one agent per role, each with its own context window, identity, and tool scope, and manage them like teammates. The unlock is management skill, not technical skill.
Position B — Agents are tools you give a goal and stay out of
- Operator: Boris Cherny
- Card: Give the model tools and a goal; do not hard-code the workflow, Underfund teams deliberately so AI substrate, not headcount, absorbs the work
- Claim: Give the model tools and a goal; do not hard-code the workflow. The right move is restraint: less scaffolding, less workflow, more abstraction. Underfund teams so substrate absorbs work; don't manage agents as headcount.
Conditions distinguishing them
- Loop type: Vo runs a 9-agent CPO operating loop where each agent has a specialised JTBD (writes PRDs, reviews dashboards, briefs stakeholders). Cherny is building a coding agent (Claude Code) where the user's job is one task and the agent figures out execution.
- Output structure: Vo's loops produce many heterogeneous artifacts that need consistent identities + tool scopes. Cherny's loops produce one homogeneous artifact (code change) where over-scaffolding limits the model's ability to discover its own path.
- Trust horizon: Vo emphasises progressive trust because each agent persists across sessions; Cherny emphasises freedom because the agent is short-lived and goal-bound.
Resolution / synthesis
Both positions converge on substrate-runs-loop / humans-run-alignment (Substrate runs the loop; humans run alignment and taste); they diverge on whether the substrate looks like a team or a tool. The reconcilable rule:
- Heterogeneous, persistent JTBD → agents-as-team (Vo): named identities, scoped contexts, progressive trust.
- Homogeneous, ephemeral, goal-bound tasks → agents-as-tools (Cherny): minimal scaffolding, give the goal, stay out.
This is genuinely orthogonal — same operator could deploy both shapes for different problems. The contradiction reveals a missing taxonomy axis: persistence + heterogeneity of the agent's job determines which shape applies.