Appendix — Radical Workflow Redesign Exercise
This appendix captures the DirkAI tutorial exercise on radical workflow redesign for 2–10× productivity gains — the conversation that surfaced the MR review bottleneck driving this playbook.
Core lesson
Organizations adopt AI/agentic tooling but see modest gains because capability is layered onto human-centric workflows instead of workflows being redesigned for agents.
| Human-centric | Agent-centric |
|---|---|
| Handoffs, calendars, tacit judgment in sequenced steps | Modular work with explicit inputs/outputs and systems of record |
| Humans as primary workers | Agents coordinate execution; humans set intent, guardrails, exceptional judgment |
| Ad hoc context reconstruction | APIs, structured context, continuous monitor→act→verify loops |
Conversation arc (summary)
| Stage | Finding |
|---|---|
| Bottleneck | Review latency — PRs wait on specific people; CI/triage automatable but review serializes delivery |
| Audit — dominant driver | Missing context in the MR → review ping-pong across days; policy ambiguity second for sensitive changes |
| Gauge — outcomes | Faster time-to-ship; fewer defects/hotfixes; protected margin and renewals; predictable delivery without trading speed for safety |
| Gauge — scoring | Medium repeatability; high business impact; medium–high complexity |
| Engineer — first shrink target | "Reconstruct context from scattered systems" before review — require machine-assembled review packet (risk, tests, owners, links) so first human touch is judgment, not archaeology |
| Takeaway | Adding AI to an old loop moves the bottleneck; leverage is workflow redesign — legible decisions, explicit policies, agents assembling evidence |
Framework recap
- Audit — Map where delays come from (routing, context, policy, load)
- Gauge — Tie changes to outcomes; score repeatability, impact, complexity
- Engineer — Choose what to eliminate or shrink first in agent-first design
Application at Soft Pyramid
Apply first to internal delivery standards (evidence packs, approvals, traceability, measurement). Extend to client-facing playbook once trustworthy and repeatable.
About the Harvard intensive
This playbook was developed during the Harvard Data Science Review 2.5 Week Agentic AI Intensive (Agentic AI: Contextualized and Applied, April 14–30, 2026), presented by the Harvard Data Science Initiative.
Author: Fakhar Khan · Soft Pyramid LLC · fakharkhan.com
Certificate verification: KKNW-EQWZ
End of The A.G.E.N.T. Playbook