Lesson 8 · 18 min
The third garden: the company, and your plan
In a company, to cultivate is to stop treating AI as software you program and start treating it as a capability the organization develops. When projects fail, the reason is rarely the model: it's a lack of context, process, and feedback. This lesson closes the course with a dated plan.
Start the lesson
In this lesson you will
- 1Choose your first process: high volume, clear rules, low cost of errors
- 2Read the symptom of a failing project and find the minimal fix
- 3Why the garden is the asset, not the model
- 4Generate your plan with dates
Goal: choose the company's first process for an agent and leave with a dated 30- or 90-day plan.
01Same model, two gardens
The model is rented, shared, replaceable. Click to compare.
A competitor can hire the same model and even the same team. They can't copy 10,000 evaluated runs. The garden is the moat; the model is not.
02From symptom to minimal fix
Why projects fail, and the smallest repair.
1. "It hallucinates"
Curate the base: less, better, with an owner. Not the whole Drive.
2. "Nobody trusts it"
Biweekly scorecard with a human reviewing 20 cases.
3. "It broke in production"
Include the ugly cases and the exceptions in the examples. And if it "did something it shouldn't have": write the list of what's forbidden and require confirmation.
4. "It stopped improving"
Ritual: every failure becomes one line of rule or example. And if it "costs more than it saves": one process, one agent, one indicator.
03Where to start, and what real cases teach
Every agent is born at level 1.
1. By area
Sales: lead qualification and research. Customer service: triage and suggested replies. Finance: reconciliation and classification of entries.
Marketing: first draft within the voice guide. Operations: data extraction and checking against a checklist. HR: résumé screening against explicit criteria.
2. Evaluation
Klarna replaced hundreds of customer service agents with AI in 2024, admitted a drop in quality in 2025, and went back to hiring humans for the complex cases: cost-cutting without evaluation is cultivation without a harvest.
"Workslop" (HBR research): almost half of workers receive AI content that looks good and is useless.
3. Culture
At Shopify, before asking to hire anyone, the team shows why AI can't do the job: the culture changes before the tool does.
Duolingo announced "AI first," got public backlash, and backed off: how you communicate cultivation matters.
4. The agent manager
The role that emerges is not a prompt engineer: it's the person who writes agent profiles, curates context, maintains examples, reads scorecards, and runs the feedback ritual.
The company that already knows how to onboard, give feedback, and measure results gets ahead.
04The company's 90-day plan
Copy it into the project document and add dates.
# First agent — <process> owner: <name> ## Days 1–30 - [ ] Process with volume, a clear rule, and low cost of error: - [ ] Agent profile written (role, result, owner, three limit lists) - [ ] Curated context (product, ideal customer, policy, glossary) - [ ] 20 annotated examples, including the ugly ones - [ ] Agent at level 1 (recommends) ## Days 31–60 - [ ] Scorecard running (sample of 20 cases reviewed by a human) - [ ] Two feedback rituals: failures become rules and examples - [ ] Cost per task and error rate measured ## Days 61–90 - [ ] Move up to level 2 if the scorecard allows it (90%+, 50+ evaluated, zero serious errors in 2 weeks) - [ ] Method documented - [ ] Second process chosen - [ ] Agent manager appointed
05The cultivation map
The course's four main deliverables and what each one leaves ready. Click an exercise to see its "done when."
Profile and the five files
2 exercises
An agent described by its eight elements, in one folder.
View the lessonMaster file and failure log
2 exercises
An agent in use every day, with FALHAS.md and a runbook.
View the lessonLog and first rule
2 exercises
The week's success rate and version v1.1.
View the lessonWritten criterion
2 exercises
Current level and the numbers to move up, in the profile.
View the lesson- 1/4
The files
Start with the profile: role in one sentence, a human owner, and the three limit lists. The other four files can start out empty.
- 2/4
Daily use
Works for Jarvis or for personal cultivation (lesson 4): one role, daily use, one line per error.
- 3/4
The cycle
Requires one week of use. Three fields per run; the failure that repeated most becomes the first rule.
- 4/4
Autonomy
Only after you have numbers. The agent stays at level 1 until the criterion is true.
Recommended order: each deliverable uses what the previous one left ready.
06Generate your plan with dates
Pick a start date and how much time you have per week. The plan respects the exercises that need days of real use.
07Final test
4 questions about the whole course.
Question 1 of 4
Which process is the best first candidate for an agent in the company?
Question 2 of 4
"It worked in the pilot, it broke in production." What is the minimum fix?
Question 3 of 4
The company swaps model X for model Y. What happens to the person who cultivated?skill
Question 4 of 4
Which sentence sums up the course?skill
Choose an option. The explanation appears after your attempt.