Structured Automation with Agentic AI
Talk at the European Space Agency (ESA): AI agents deliver productivity gains on API-driven workflows. Production use requires guardrails and careful prompting.
AI agents promise autonomy — and deliver it, just not always in the direction you had in mind. This talk draws the balance from three months of intensive practice with agentic tooling: Claude Code, Codex, Gemini CLI and local Qwen coder models, applied to real community operations — ticket allocation, media production, reporting, and the many small process chains that otherwise cost volunteer conference organisers their weekends.
The central observation: productivity gains happen where workflows run over documented APIs and the agent stays on a short leash. Where either is missing, AI-assisted engineering tips into what the talk calls "augmented arrogance" — impressive-looking results on inaccurate data that maintainers pay for later. Between vibe coding and dependable automation lies structure: narrow use cases, guardrails, frequent commits, git worktrees, subagents — and honest expectation management.
The talk combines lived experience with perspective: why a language model remains a parrot, even in a flock. Why more data and bigger models no longer drive quality. And why one effect outweighs the rest: AI enables the generalist.
Key theses
- Productivity gains come from documented API workflows — not from open-ended exploration tasks.
- Keep agents on a short leash: narrow use cases, guardrails, only the necessary tools, commit often.
- Git worktrees and subagents parallelise agentic work without losing control.
- Skipping the agent's output doesn't save thinking — it defers it, at a higher price.
- Scaling via data volume and model size has stalled; the leverage is in structure and steering.
Context
The talk grew out of automation practice in the European Python community and was delivered at the European Space Agency (ESA) in Darmstadt, among other venues. The slides are deliberately designed as a visual backdrop — the substance is in the delivery.
