Our working answer: agents stop being features and start being staff. The impact isn't a smarter search box bolted onto old software — it's business processes where the agent carries the judgment steps: reading the situation, making the mid-level call, executing through deterministic tools, and writing down what it did. We call these agentic workflows, and this lab has been building, breaking, and shipping them since 2025.
The experiment line
| Experiment | What it proved | Status |
|---|---|---|
| Sales Intelligence | The benchmark, 2025. A directive-driven agent ran a full multi-step revenue workflow — research, scoring, reporting, outreach drafting — reliably enough to trust. Everything since evolves this pattern. | Benchmark |
| The Foundry | One editorial engine, many publications. A control plane where personas, sources, memory, and calendars are data — and interchangeable runners execute the work on schedule. Currently producing a weekly professional column, unattended. | Live |
| QuickLaunch (day job) | The research deployed commercially. Agentic workflow patterns from this lab ship inside QuickLaunch, where Todd leads AI: the Insight Agent is in active development for an October release, and the Agentic Design Studio's patterns and prototypes come from Clark — not from Figma. | Shipping · October |
What we've learned so far
- Push judgment up, push execution down. Agents decide; deterministic scripts do. Reliability comes from keeping those layers separate.
- Green checkmarks lie. The operational failures that hurt were the silent ones — jobs that "succeeded" while doing nothing. Critical paths now fail loudly or not at all.
- The org chart becomes a cast. When agents are staff, you manage them like staff: track records, corrections, standards — and occasionally, firings.