Business Agents 101 · curriculum

Six modules. From AI chat to building agents.

For analysts, product people, ops, leaders. Hands-on from module one. Your own work, your own data, your own systems. No code. Plus two optional extensions when a cohort wants to go further.

Format
Building is thinking. Every concept emerges from doing it first — never explained before the exercise that demonstrates it.
Who it’s for
Analysts, product people, ops, leaders. Anyone who works with process and has a real challenge to bring into the room.
What you need
A laptop. A Claude account. A real work challenge. About 45 minutes of prework. No coding, no prior agent experience.
Modules 1–3

Build, break, understand.

First working agent by module one. A multi-agent system by module three. One challenge that keeps growing — not throwaway exercises.

I.

Getting Going

With the right guardrails, the output is genuinely yours — not generic.

What you build
A personal artifact from your own LinkedIn profile. Not copy-paste — something differentiated enough that only you could have produced it, using positioning frameworks as guardrails.
Teaching moment
Structured guardrails produce higher quality than unstructured prompting. You catch your first fabrication by using domain knowledge as the evaluator.
II.

Building Agent Systems

A system remembers, grows, and compounds. Chat doesn’t.

What you build
An LLM brain on your next big work challenge — curated from your wiki, recent work, and practitioner reading. A custom agent, written in plain markdown, that does real work against that brain. Plan-mode reviews before multi-step actions run.
Teaching moment
Three layers: sources, brain, rules. You see why plain text beats a database here — and whether the knowledge base is compounding or merely growing.
III.

Multi-Agent Systems

Hire three agents to search. Three more to decide. The filesystem is the meeting room.

What you build
Three retrievers — one for your wiki, one for your collab suite, one for curated internet — and a synthesizer that reconciles what they find. A real strategic answer produced live, against your own challenge.
Teaching moment
When splitting earns its keep, when it doesn’t, and when one good agent with a good prompt wins. Handoff failure modes named, not hidden.
Modules 4–6

Discipline.

You built something that works. Now make it something you can trust. Security, quality, and evaluations as infrastructure — not theatre.

IV.

Security

You can’t see safety in the output. You need a way to check — and the practice is running the check.

What you build
Two checks plug into your system: your company’s own policy rules and a generic agent-security skill. You audit the agent you built in Module 3 against both. Apply a mitigation. Observe what’s left.
Teaching moment
“I can’t tell” is a real answer, not a failure. The best move is sometimes to not open the door — naming that is part of the discipline.
V.

Output Quality & Hallucination Control

Quality is measured, not assumed. Run the bake-off.

What you build
Four detection methods run in parallel against a gold standard you wrote in two minutes. A meta-evaluator scores them. The winner is saved as a reusable judge with a stated scope and a named known-limit.
Teaching moment
Method selection is empirical, not authoritative. You pick the winner by reading the scores. What a judge can and can’t reach matters as much as what it catches.
VI.

Evaluations

The eval is infrastructure, not an artifact. Two loops compound.

What you build
A self-improving eval loop — generations run in parallel, the judge scores them, a meta-agent edits the judge’s rules when it misses something, and the rule-change history proves it’s actually learning.
Teaching moment
Eval as infrastructure vs. eval as artifact. Generation sharpens under eval pressure; the eval sharpens from watching its own misses. On disk, re-runnable, sharper each cycle.
Modules 7–8 · optional

Two extensions, when you want the flywheel.

The six-module cohort stands on its own. When a room has appetite for more — sharing agents across teams, or the move that changes the arithmetic — these two modules extend the program.

VII.

From Personal to Team

You can’t really share an agent. You can share context, a skill, the output, or an interface — and the choice is a strategy decision, not a deployment decision.

What you build
A jobs-to-be-done interview with your teammate’s actual job. A sharing strategy matched to the outcome they want moved. Both a technical plan and a people plan. A pre-mortem on the social failure you’re not yet seeing.
Teaching moment
Three walls every practitioner hits: the absorption bottleneck (learning-rate cap), the access-trust gap (access is easy; production trust is scarce), and discoverability — nobody uses what they can’t find in the flow of work.
VIII.

Agents Building Agents — The Flywheel

The tool that builds tools compounds.

What you build
An agent that generates another agent. Then the room runs a strategy synthesis together, using three disciplines each participant applies with their own agent: the crux skill (Rumelt) names the load-bearing obstacle; the assumption-test skill (Roger Martin) ranks what would have to be true; the pre-mortem skill (Klein/Kahneman) names how it fails. The synthesised AI strategy is the artefact.
Teaching moment
The three thinking disciplines stay with the participants — working versions go home. Strategy emerges from the system the room just built — not the other way around.
Outcomes

Real agents. Real work. Still running.

A handful of what participants have shipped — each now doing the work it was built for.

A people-ops lead

A five-level framework for AI ways-of-working (L1 to L5), with an eleven-question assessment, individual and manager dashboards, and cognitive-load measurement. Planned integration with the L&D platform and team OKRs. Queued for company-wide rollout.

A strategy analyst

A living strategy vision in Confluence and a continuously-updated valuation model. Answers valuation questions in minutes that previously took a week from external advisors. The “AI-first strategy process” the company had been chasing for a year, finally running.

IV. How we teach

Six rules the curriculum obeys.

Building is thinking

Concepts emerge from doing. We never explain something before the exercise that demonstrates it.

Your data, not toy data

Your profile, your policies, your domain. Not sample_data.csv.

Show the failure mode

For every capability, you see what goes wrong without the discipline. The violation. The undetected fabrication.

Your reality replaces our defaults

Security policies, evaluation criteria, domain knowledge. Clearly marked places where your organization plugs in.

One new thing at a time

Each exercise adds one concept. Never two at once.

The agents stay with the company

Every agent the cohort builds runs on your data, in your environment. Owned by your company after the cohort closes.

Ready when you are

Bring a challenge. Leave with a working agent.

Tell us the audience and the rough timing. A 30-minute scoping call is enough — we bring the shape of the engagement and confirm the price.

Start a conversation → See pricing →

Helsinki · Remote & on-site