Shift Units from Task Agents to Digital Workers

Task agents handle discrete outputs in one or few sessions, like Claude Code writing over hours or Deep Agents managing 30 tool calls without unraveling. They rely on scaffolding: planning tools, subagent isolation, filesystem scratchpads, context discipline, and production evaluation. Digital workers, however, manage recurring outcomes over long horizons, like ongoing functions in a business. You onboard them with identity contracts (name, one-sentence purpose, explicit scope/non-scope, single human sponsor) rather than deploy them. This resolves issues like ownership disputes, audit permissions, memory versioning for thesis updates, and governance logs—problems beyond the model or task loop, demanding a harness layer above.

Build the 6D Harness to Answer Hiring Questions

Mirror organizational hiring by engineering six dimensions on one canvas:

  • Identity: Contract defining who (name, purpose, scope, sponsor) for hireability and retireability.
  • Context: Inherit from the function—strategy, brand voice, key humans, systems—not invented by engineers, preserving separation between function and software project.
  • Capability: Named units with triggers, outputs, acceptance criteria; compose, validate, fix independently like existing harness skills, but accountable per boundary.
  • Conduct: Autonomy as gradient per action risk/maturity (e.g., detect funding autonomously fast, outreach needs indefinite human approval); engineer predictability for surprises so sponsors anticipate behavior.
  • Cognition: Hierarchy of truth sources, conflict resolution, closed learning loops against observable outcomes (e.g., calibrate investment scoring on past passes, tune capacity proposals from decision divergences).
  • Governance: Named owner, KPI, contingency; dedicated service account, action audit trails (worker vs. human), retirement path, sponsor-pull pause—no engineering on-call needed.

This treats stable hiring questions as engineering surface, transferring task harness discipline upward.

Deploy in Production with Phased Autonomy

In a thousands-person services firm, one worker owns capacity-management: 5 capabilities, 1 sponsor, 3 autonomy phases (propose/approve Phase 0 → approve-by-exception Phase 1 → execute-in-envelope Phase 2) progressed by metrics over 12 weeks. 6D canvas gates production.

In a VC firm, 5 workers handle investment pipeline, portfolio stewardship, capital raising, reporting, decision intelligence—each with sponsor, KPIs, progression. Sequence waves: first capability to Phase 0 before next worker builds. Same canvas scales across functions/firms.

The product is the function (recurring outcome once human-owned), with digital worker alongside human—prompt engineering scaled to harness, now to wider harness for production reality.