The Agentic Operating System — workshop build

Most teams meeting agentic AI for the first time get stuck on one of three blockers: tool design, orchestration choice, and the gap between a working demo and a system that survives Monday morning. The AIMED workshop format compresses the answers into one day of hands-on building.

• 1Engineers had used Claude and Copilot but never built an agent loop end-to-end
• 2Multi-agent orchestration is widely-discussed but rarely walked through in code
• 3Workshops that just demo finished agents leave attendees without the muscle memory to ship their own
• 4A real artefact that survives the day is the difference between a workshop and a memorable workshop

A day-long live build of "the Agentic Operating System" — a multi-agent shell with a supervisor (planning, decomposition), handoff agents (parallel reads, sequenced writes), shared tool registry via MCP, and observability wired in from line one. Every attendee leaves with a running shell on their own laptop, the source, and the patterns to extend it.

The shell is generic — file reads, shell commands, web search. By the afternoon, attendees were forking it to wire in their own tools: GitHub API tools, Jira tools, internal codebase searchers, and so on.

• Files MCP: read/write scoped to a workshop sandbox dir
• Shell MCP: scoped to a small allowlist of commands
• Search MCP: web + GitHub code search wrapped behind one tool
• Each attendee added one tool of their own by the end of the day

The single most-fed-back lesson from the day was that wiring Langfuse before the first multi-agent call (not after) changed how attendees debugged. Once the per-step trace was visible, problems that took minutes to find took seconds.

• Per-step trace with the workshop attendee's name as session prefix
• Cost-per-attendee surface — running cost visible all day
• Cache-hit rate climbed from 0% at hour 1 to ~72% by hour 6 as the system prompt stabilised

A running shell on their own laptop. The source on GitHub. Three working MCP servers. A supervisor + handoff pattern they can extend. And — most valuable — the muscle memory of having built this end-to-end.

• 01Workshop attendees retain ten times more from one full build than from ten finished demos. Build it together; do not show it pre-built.
• 02Langfuse-from-line-one was the highest-leverage decision in the curriculum. Without per-step trace, hour 5 onwards becomes guesswork; with it, debugging is direct.
• 03Three specialists with narrow scopes beat one big agent loaded with everything — confirmed live by an A/B at hour 6, with the audience watching the eval-score deltas.
• 04The "add-your-own-tool" hour is what makes the workshop generative. The variety of extensions in one room is what the attendees remember three months later.

The Agentic OS shell is now the curriculum spine of the corporate training programs we deliver. Customers come in with their own integrations to bolt on, and the same 8-hour structure compresses neatly into one full day of an 8-day intensive.

• Same pattern delivered as Day 5 of the corporate 8-day intensive (IT track)
• The Agentic OS GitHub repo is the public reference implementation
• Workshop graduates often return with their own MCP servers and extensions to share

Want this workshop for your team?

The same format runs on-site for engineering teams of 8–14. We customise the integrations to your stack — by the end of the day your team has shipped a working agentic shell in your codebase.