Topics
-
From anti-AI to an autonomous workflow
Before 2026 I was anti-AI, sure that no machine would tell me how to write code I had to think through myself. Now an autonomous workflow delivers what's needed.
-
Think first, then delegate
The nuances of coding with AI agents. The best results come when a person thinks first, then delegates and talks the work through with AI.
-
AI failure modes and dedicated auditors
Where AI agents commonly go wrong, and how dedicated auditors came out of it. The fidelity auditor was added after an AI silently cut a 491-line spec to 122 lines.
-
Gates agents can't talk their way past
Deterministic gates, like custom static-analysis and dependency rules or shrink-only baselines, and how an agent games a gate that measures only one thing. Behind such gates, a 15-day refactor went through 51 pull requests with no post-merge reverts.
-
A/B testing models on real work
Blind A/B tests on real work items, with hidden tests and blind cross-vendor judges. It's how I pick which model implements and which one reviews.
-
Getting started with Claude Design
How I use Claude Design and how to start with it. Its screens reach coding agents through my open-source Claude Design Playbook.
-
Contracts for repeatable work
A contract is a set of written rules an agent must follow. Anything repeatable in my workflow becomes one, for example for tests or for writing a plan.
-
Tests an AI can't fake
Every test has to fail against deliberately broken code first. A meaningless test is worse than none.
-
MCP for AI work, and MCP vs CLI
Using MCP to support the work AI agents do, and how it compares with a CLI. My orchestration runs on a PostgreSQL MCP server.
-
How I stopped compacting
State lives outside the session: MCP stores what other agents discovered, and tracker tasks are the shared source of truth. Every session starts from there, so compaction isn't needed (Agentic Jams, 17.06.2026).
-
Specialist agents
Agents built for one job each, like implementers, reviewers and auditors, and how they fit together.
-
Setting up autonomous orchestration
Creating and setting up a workflow that orchestrates itself. In mine, agents supervise agents and escalate by my rules.
-
Parallel agents on worktrees
Running several agents at once on separate worktrees without them corrupting each other's state.
-
Temporary workflows for one-off jobs
Setting up a short-lived workflow for a single need. I did it to migrate Twig macros to components.
Past talks
- rg-dev Meetup #67, 23.09.2026: From Directing Agents to Autonomous Delivery: mini-deliver — One Skill, the Whole Loop
- Agentic Jams (Rzeszów, first edition), 17.06.2026: How I Stopped Compacting — MCP as durable memory for an AI delivery pipeline
- rg-dev Meetup #66, 27.05.2026: From Prompting to Directing Agents: Practical AI-Native Development
- Xebia Internal Conference, 11.2023: GitHub Actions & Continuous Delivery on Shared Hosting