The frontier firm: agents as digital labour, humans as agent bosses
Org design for human + agent teams: ratios, decision rights, gates and the unit economics
The question: What does an organisation look like when some of its "workers" are agents — and how do you design it so the humans stay accountable and the economics hold?
Microsoft's 2025 Work Trend Index named the "frontier firm": human-led, agent-operated teams. The pattern is arriving in functions now. Designed well, it is Lesson 22's governed automation at organisational scale; designed badly, it is an org chart with nobody accountable for the seams.
What the lesson covers
The frontier-firm pattern (Microsoft WTI 2025): work is organised around outcomes; agents execute bounded tasks; humans set intent, supervise, handle exceptions and own results — "agent bosses". The 2025 evidence in the report and elsewhere shows firms moving from pilots to a first wave of agents in real functions, with human:agent ratios becoming a design variable rather than a metaphor.
Design it with the course's own tools. Each agent is a six-box automation (Lesson 22) on an autonomy rung (Lesson 10) with a human gate on the irreversible. **Decision rights** are explicit: which outcomes a human owns, which actions an agent may take alone, what escalates. **Ratios** follow the work: how many bounded agent tasks one accountable human can supervise depends on reversibility and volume, not on ambition. **Reach** is scoped: every agent's tools and tokens at least privilege (Lesson 33).
The seams are where it fails. An org chart that shows agents but not who owns the handoff between an agent's output and a human's decision is the Lesson 16 failure in new clothes: capability sitting unused or, worse, acting unaccountably. Name the owner of every seam; give them the log, the heartbeat and the authority to stop the agent.
Cost it honestly with Lesson 31: agents carry inference, review, and maintenance lines; "digital labour" that is cheap per task can be expensive at volume with poor routing, and the human supervision is a real line. Cost per completed outcome, fully loaded, against the human-only baseline — and a sensitivity at triple volume.
Culture last, because it decides adoption: people supervise agents well when they are measured on outcomes and given the authority to override, and badly when they are measured on throughput and told to trust the machine. The human-change discipline of Lesson 32 applies — small, instrumented, decided in public.
Key points
- Frontier firm: outcomes-first, agents execute bounded tasks, humans set intent and own results (WTI 2025).
- Each agent is a six-box automation on an autonomy rung with least-privilege reach.
- Name the owner of every seam between agent output and human decision — that is where it fails.
- Cost per completed outcome, fully loaded, with a volume sensitivity; human supervision is a real line.
- Measure supervisors on outcomes and give them override authority; roll out in ninety-day steps.
Framework — Outcome → Agents (six boxes, rung, reach) → Human owner of every seam → Unit economics → Decide in public
Organise around outcomes; each agent governed; every seam owned; costed per completed outcome; rolled out in ninety-day steps.
The lab
Design a human + agent organisation for one function and cost it.
Open this lesson, its lab and its quiz
Sources and further reading
- Work Trend Index 2025: The Year the Frontier Firm Is Born — Microsoft
- Building Effective Agents — Anthropic
- Lesson 22 — trigger, AI step, validate, gate, act, log — Applied AI Academy