Agentic Software development with Claude Code
Agentic Software development with Claude Code
5 May 2027, 09h30 - 17h00
General context
Coding agents such as Claude Code can plan, write, test, and refactor software on their own. That makes typing cheap, but the decisions, the checks, and the understanding still have to stay with you. This hands-on course is for researchers with a programming background who turn research questions into libraries, tools, or pipelines.
In one day, you take a project from design to a versioned release, following a plan-do-check-act loop in five modules: design and structure, the agent and its context, build and verify, review and secure, and ship, track, and evolve. You work on your own repository or research problem, or a ready-made project, using Claude Code in depth together with Git, Docker, CI/CD, testing, and experiment tracking. Every module ends with a decision you make yourself, a step without the agent, an independent check. The way of working carries over to other coding agents (like opencode, PI harness, etc.)
- Plan a software project with a coding agent: write a brief, set up a clear repository structure, and record design choices as Architecture Decision Records (ADRs).
- Configure a coding agent for your project with instructions, skills, permissions, and hooks, and manage its context deliberately.
- Build and verify: let the agent work in small, reviewable steps from specifications, and check the result with tests it cannot see, including property-based and mutation testing.
- Review and secure agent-written changes with separate checks for bugs, simplicity, and security, and run these checks automatically in CI.
- Ship a reproducible, versioned release with containers, CI/CD, and tracked experiments, and explain and defend every change in your repository.
- Each module opens with a short conceptual introduction and a live demo, followed by guided exercises.
- Bring your own repository or research problem; a ready-made project is available if you don't have one.
- You work in pairs; roles swap per module.
- Each module includes a decision you make yourself, a step without the agent, and an independent check.
- You leave with a working repository and an agent setup you can keep using.
- Comfortable with Python programming; programming basics are not covered.
- Working knowledge of Git (branches, commits, pull requests) and the command line.
- Basic familiarity with Docker and CI/CD is an advantage.
- No prior experience with coding agents is needed.
- Bring a laptop, a GitHub or GitLab account, and, ideally, your own repository or research problem.
- Coding agent: Claude Code: plan mode, CLAUDE.md, skills, hooks, permissions, subagents, workflows, /code-review, /simplify, /security-review, headless mode
- Version control, CI/CD: Git, GitHub or GitLab, GitHub Actions or GitLab CI
- Python tooling: uv, ruff, mypy or pyright
- Testing and validation: pytest, pytest-bdd, Hypothesis, mutmut, pandera, pydantic
- Containers: Docker, devcontainers
- Security: gitleaks, pip-audit, Dependabot
- Tracking and pipelines: MLflow or TensorBoard; Argo Workflows, Nextflow or Snakemake
- Formal verification: Lean or other systems (short demo)
AI methods used
• Agentic coding with large language models: an agent built on Anthropic's Claude models plans a task, reads and edits files, and runs commands and tests in a loop, with a human approving each step.
• Context engineering: project instructions, skills that load on demand, and active management of what the model sees.
• Multi-agent patterns: subagents for independent search and review; orchestrated workflows that run several agents with cross-checks.
• LLM-based code review: separate automated checks for bugs, simplicity, and security vulnerabilities.
• Agents in automation: Claude Code running headless in a CI/CD pipeline.
• Not covered: training or fine-tuning models.
Trainers
Jens Krijgsman
Jens Krijgsman leads the AI Lab at Howest University of Applied Sciences in Kortrijk, Belgium. His work centers on agentic coding and applied LLM solutions, backed by hands-on experience with data and ML pipelines, automation, and deployment. He also teaches MLOps at Howest and is involved in international collaborations.
Program
Welcome, set-up check, the plan-do-check-act loop
Module 1 · Design and structure: brief, repository template, ADRs
Coffee break (provided)
Module 2 · The agent and its context: instructions, skills, context management, guardrails
Lunch (provided)
Module 3 · Build and verify: specifications, small steps, held-out and property-based tests
Module 4 · Review and secure: bug, simplification and security review; CI on every pull request
Coffee break (provided)
Module 5 · Ship, track and evolve: containers, versioned release, experiment tracking
Wrap-up: peer defence, what transfers to other agents, Q&A
Practical info
05 May 2027
Ghent - VIB/UGent FSVM II
Technologiepark 75
9052 Zwijnaarde
Belgium
05 May 2027
Ghent - VIB/UGent FSVM II
From Ghent Sint-Pieters station, you can take a bus to Technologiepark. Please check Routeplanner De Lijn for schedules.
05 May 2027
Ghent - VIB/UGent FSVM II
Looking for a custom cycling route? Try the online cycling route planner to easily and quickly identify the fastest, safest and/or most scenic route to any destination in Ghent.
Cycling route planner
Shared bicycles:
Several providers are active in the area.
Take a look at their platforms:
- Indigo Weel Pro app
- Bluebike
- DOTT
- BOLT eBikes
- Donkey Republic
There are also some public bicycle pumps at the park.
05 May 2027
Ghent - VIB/UGent FSVM II
There is only one entrance to Technologiepark. At the entrance, please take a ticket - Parking is only allowed in regular parking spots (for instance in front of the building) and in the parking tower. Parking alongside the roads or in other places where there is no regular parking is prohibited.