COTRUGLI Business School · Dr. Tali Režun
Agentic Engineering
The Centaur Developer Programme
Recognised by
From copy-paste coding to orchestrated, context-engineered agent teams
Less talking, more building. Every concept in this programme is learned by doing it on a real repository.
- Audience
- Software developers, tech leads, CTOs and engineering managers at technology companies; technical founders and independent developers
- Starting point
- Teams that have tried ChatGPT or Claude in a browser, copied code into VS Code and found the process slow, manual and hard to review
- Duration
- 4 modules of 2 to 5 hours each: 8 to 20 contact hours. Practice between modules is expected
- Format
- Hands-on workshop. Live, hybrid or online
- Languages
- English, Slovenian or Croatian
- Primary harness
- Claude Code, with OpenCode as the open, multi-model harness. Codex or another harness on request
- Final outcome
- An agent-ready engineering setup on a real codebase from your own work, plus a capstone project
- Offered by
- COTRUGLI Business School (corporate and open enrolment) or directly by Dr. Tali Režun
Why it exists
Most teams have already tried AI: a developer asks a chat window for a function, pastes it into the IDE, finds it does not fit, and asks again. The process is manual, context is lost at every step and progress is hard to measure. The tools are not the problem. The workflow is. Agentic engineering replaces the chat-and-paste loop:
- Agents work inside the repository: they read the codebase, run the tests, use git and call real tools.
- Context is engineered, not improvised: foundational documents, standing briefs, handoffs and a persistent context engine.
- Work is orchestrated: a lead agent plans and delegates to sub-agents, routing simple work to cheap models and hard work to frontier models.
- Everything is verified and versioned: tests, browser automation, pull requests and scheduled reviews.
The Centaur thesis
In freestyle chess, human-plus-machine teams beat both grandmasters and the strongest engines playing alone. The human contributes specification, judgement, taste and context; the agents contribute implementation speed, breadth and stamina. The goal is not to write less code. It is to ship more of the right software, with fewer surprises, at a lower cost per feature.
Learning outcomes
By the end, you will be able to:
- Choose and configure an agent harness (Claude Code, OpenCode, Codex) and explain why the harness, not only the model, decides the quality of the result.
- Decide between cloud, open-weight and local models, weighing capability, cost, data sovereignty and hardware.
- Set up a project for agent work from day one: foundational docs, a standing brief, skills and MCP configuration.
- Engineer context across sessions and harnesses with the three context layers, the 80% rule and handoff files.
- Build and connect MCP servers and skills well enough to debug them.
- Orchestrate agents: plan-then-execute, sub-agents, model routing, parallel work and verification loops.
- Automate engineering oversight with scheduled tasks connected to GitHub: PR review, daily build reports and docs-drift checks.
- Connect agents across harnesses with AgentLink, and build a custom OpenCode agent on open-weight models.
- Treat documentation as infrastructure and deploy with scoped credentials.
Four modules
Every module follows one rhythm: a short framing, a live demonstration on a real repository, then hands-on work on your own machine. Ten assignments, E1 to E10, lead to a capstone.
Foundations & the Harness
Stop pasting code. Put the agent inside the repository, and learn MCP by building one.
- The Centaur thesis, and three eras of AI coding: autocomplete, CLI agents, orchestration
- Anatomy of an agent harness and the harness landscape
- Cloud versus local models, subscription economics and token discipline
- Standing briefs: CLAUDE.md, AGENTS.md, GEMINI.md
- Model Context Protocol, learned by wiring real tools into an agent
- E1Workstation & accountsOutput: A verified, working agentic workstation
- E2Make a repository agent-readyOutput: CLAUDE.md and AGENTS.md committed to the repo, plus a one-page harness comparison
- E3Build the Atlas agent and learn MCP by doingOutput: A governed Atlas agent with working MCP tools
Context Engineering
The master skill. Keep a large build coherent across sessions, machines, models and harnesses.
See the workflow- Prompt, context and harness engineering; the three-phase plan
- Foundational documentation, and reverse-engineering docs for an existing codebase
- Three kinds of context, the 80% rule and handoff files
- The Curator as a context engine, readable from any harness
- Multi-harness development and AgentLink, harness-to-harness communication
- E4The foundations packageOutput: A docs/ folder that grounds any new session, in any harness
- E5Install The Curator as your context engineOutput: A second brain plus portable working state for one real project, verified across two harnesses
- E6Harness relay with AgentLinkOutput: A working cross-harness channel and a team policy for agent-to-agent work
Orchestration, Skills & Automation
One developer, a team of agents, and an assistant that watches the repo while you sleep.
See the workflow- Plan-then-execute, sub-agents and model routing
- Parallel work with git worktrees; skills, hooks and custom commands
- Verification loops with Playwright: agents do not fail loudly, they fail confidently
- Scheduled tasks on GitHub: daily build reports and PR review for CTOs
- OpenCode as a build-your-own harness on open-weight models
- E7An orchestrated feature buildOutput: A merged or reviewable PR built through orchestration, a sub-agent set and one team skill
- E8Engineering oversight with scheduled tasksOutput: A daily build report and an automated PR reviewer running in the cloud
- E9Your own agent in OpenCode on open-weight modelsOutput: A custom agent, a mixed-fleet configuration and a cost/quality comparison
Documentation, Infrastructure & Capstone
Docs that keep themselves current, infrastructure agents can reach safely, and a capstone that pays off on Monday.
- Documentation as infrastructure; READMEs for open and closed source
- GitBook versus GitHub docs, synced automatically
- Cloudflare, Vercel, Firebase, Google Cloud and DigitalOcean for agentic teams
- Scoped access and a multi-model security audit before release
- A 30-day adoption plan for a team
- E10Documentation pipeline and agent-safe deployOutput: A live deployment, self-updating documentation and a docs-drift check
Capstone: your agentic engineering setup on real work
Chosen in Module 1, built through Modules 2 to 4 and presented in Module 4 in 10 minutes. Deliver a foundations package, standing brief, context engine, at least one sub-agent or skill, one scheduled task and a measured before/after.
Output: A working setup on real work, documented so a colleague can adopt it
Delivery formats
The four modules are fixed; their length is not. Delivery can be on consecutive days or spread over several weeks, live on-site, online or hybrid.
Adapted per client: harness, codebase, git platform, infrastructure, and a local-first track on open-weight models for data sovereignty.
Before you join
- A capable laptop (16 GB RAM minimum) and administrator rights. Missing admin rights are the most common reason a participant falls behind.
- A frontier subscription, such as Anthropic Claude Pro (minimum) or Max (recommended), or an OpenAI plan with Codex.
- A GitHub account and a free Cloudflare account.
- Working knowledge of git, the terminal and one programming language. No prior experience with agents is needed.
Tools in the programme
Built on tools Tali develops and runs in production. Participants keep access to all of them.
Who teaches
Dr. Tali Režun, Vice Dean for Frontier Technologies at COTRUGLI Business School, serial entrepreneur and academic working at the frontier of AI and Web3. He designed the Chasing Jarvis programme for the Vanguard MBA. He teaches what he builds: the last year's products, including The Curator, Lumina and the MountVacation MCP server, were shipped by orchestrating coding agents.
Bring it to your team
Tell me about your team and stack. To be agreed before scheduling: format, delivery, harness, participants, subscriptions, IT clearance, the codebase and data guideline, and the capstone scope.