Know when to build a workflow, an agent or a team of agents.
Agentic AI Buildcamp for Engineers & PMs. A four week buildcamp on Maven where you ship two production systems, each measured by the outcome it delivers, plus your own AI employee. 16 November to 11 December 2026, $997.
Workflows16 to 20 November
The Map and the OutcomeLive, 17:00 UTC
Workflows, the Regulated PipelineLive, 17:00 UTC
Office hoursDue: A regulated workflow
Agents with tools, memory and permissions23 to 27 November
Agents with Tools, Memory and PermissionsLive, 17:00 UTC
The Board and the Support ResolverLive, 17:00 UTC
Office hoursDue: An agent with tools, memory and permissions
Agentic systems30 November to 4 December
Agentic Systems, Topologies and TriggersLive, 17:00 UTC
Durable Work That Survives a CrashLive, 17:00 UTC
Office hoursDue: A team of agents that survives a crash
Guardrails, evals and a second tenant7 to 11 December
Guardrails, Policy and Cost per OutcomeLive, 17:00 UTC
Evals, a Model Swap and a Second TenantLive, 17:00 UTC
Demo dayDue: Guardrails and evals you can showDemo: Your own AI employee
Why do teams build an agent when a workflow would do?
Most agent projects fail on the first decision, long before the code.
The shape is picked by fashion. A team of agents gets built for a job a fixed pipeline could run, and every run costs more than it should.
Nobody wrote the outcome down. Without a metric and a must never, nobody can say whether the agent works.
It falls over in production. A crash writes the same row twice, an email hides a prompt injection, and a model swap quietly makes it worse.
The buildcamp makes the shape decision first, then builds each shape on a real system, so you see what each one costs and where it breaks.
What you ship in four weeks
A regulated workflow.
The first project is a drug maker's inquiry pipeline where code decides every step and an eval gate fails the release on a missed adverse event.
An agent with tools, memory and permissions.
An analyst that answers only from sources the engineer may open, and a support resolver whose refunds above a limit wait for a human on Telegram.
A team of agents that survives a crash.
A manager that blocks self approval, a researcher that splits work across subagents, and a finance clerk that resumes without writing the same ledger row twice.
Guardrails and evals you can show.
Prompt injection blocked, approvals as policy, cost per outcome on the board, evals built from traces and a second tenant that cannot see the first.
Your own AI employee.
One job from your work, with its outcome, metric and must never, demoed and defended on demo day.
Start with a workflow, and earn the agent.
How the four weeks run
Two build sessions a week on Mondays and Wednesdays at 17:00 UTC, recorded in case you miss one. Fridays are live questions, where each week's project is due.
Set up before session one
Clone the starter repo, run it with your own API key and pick one job at work you would hire an AI employee to own.
Build every week
Two sessions on Monday and Wednesday, then the week's project at Friday office hours on 20 November, 27 November and 4 December.
Demo and defend
On 11 December you demo your AI employee and get feedback from Param and the cohort.
The Map and the OutcomeMonday 16 November, 17:00 to 19:00 UTC
Workflow, agent or team of agents: who decides the next step, what starts the system, how long it lives and what each shape costs in tokens. You write the outcome, metric and must never for your own AI employee.
Workflows, the Regulated PipelineWednesday 18 November, 17:00 to 19:00 UTC
Chain, route, run in parallel and check, on a drug maker's inquiry pipeline where code decides every step and an eval gate fails the release on a missed adverse event.
Agents with Tools, Memory and PermissionsMonday 23 November, 17:00 to 19:00 UTC
The model now picks the next step. The analyst answers an engineer's question only from sources that engineer may open, and never cites what it did not read.
The Board and the Support ResolverWednesday 25 November, 17:00 to 19:00 UTC
An agent that works the support inbox: it notifies, asks and waits for review, refunds above a limit only after a human approves on Telegram, and reports its cost per resolution on the board.
Agentic Systems, Topologies and TriggersMonday 30 November, 17:00 to 19:00 UTC
Supervisor, handoff, orchestrator workers and maker checker. The manager hands out work and blocks self approval, and the researcher splits a cited account brief across subagents.
Durable Work That Survives a CrashWednesday 2 December, 17:00 to 19:00 UTC
A finance clerk matches payments to invoices on a nightly schedule. We kill it partway through a run and prove it resumes without writing the same ledger row twice.
Guardrails, Policy and Cost per OutcomeMonday 7 December, 17:00 to 19:00 UTC
Block a prompt injection hidden in an email, write approval rules as policy, route each step to the cheapest model that passes, and read cost per outcome off the board.
Evals, a Model Swap and a Second TenantWednesday 9 December, 17:00 to 19:00 UTC
Evals built from real traces, a model swap the release gate blocks when it makes an employee worse, a second tenant that cannot see the first, and the rehearsal for demo day.
“You learn a system by building it, breaking it and fixing it. So that is what every workshop does.”
- Building software since 2011.
- Architect of the pivot behind Jobbatical's €11.6M Series A.
- Led the engineering work behind ISO 27001 certification on a regulated platform.
Questions about the buildcamp
Who is it for?
Engineers, engineering leaders and PMs who can read and run code and need to decide what to build with agents.
Who is it not for?
Anyone who cannot give it about two evenings a week for four weeks, or who wants a course to watch without building.
What if I miss a session?
Every build session is recorded on Maven.
Which tools and models does it use?
Each system names the model in one place, and the final session swaps it behind a release gate.
Do I keep the code?
Yes. The starter, the reference systems and your own build are yours.
What does it cost?
$997, paid once on Maven, the same worldwide.
Can my company pay?
Yes. Maven gives 20% off for 2 to 9 seats and 25% off for 10 or more. The team kit on /for-teams has a reimbursement email and an invoice in your company's name.
Do the workshops count toward it?
A workshop fee is credited to the buildcamp if you enrol within 60 days.
Where do the lessons go after it ends?
The sessions join Pro after 11 December.
Is there a cohort on coding agents?
The AI-Native Software Factory Buildcamp is in session now. Its next cohort runs on kaas.team. See /cohort/agentic-software-factory.
What is settled before you enrol
One payment of $997 on Maven, the same worldwide.
Every session is recorded.
The code you build is yours to keep and use at work.