Testing what models can do inside real systems
Agents, automation and developer tools, built to hold up outside a demo
- Project
- AI systems
- Period
- 2022–Now
- Role
- Engineering and experiments
Context
Models went from autocomplete to colleagues in four years. I adopted each step early, then moved it into how a production team actually works.
What I did
- 2022, use. Copilot and chat models: AI helps me type and think faster.
- 2023, build. Models inside software: GPT-4, vector search, the first retrieval experiments.
- 2024, architect. Retrieval, tool calling and evaluation as parts of the architecture, not a sidebar.
- 2025, delegate. Coding agents doing multi-file work, and agentic workflows adopted across the company.
- 2026, orchestrate. Many agents and models in one system: specs, skills, hooks, worktrees, evals and automated review.
The questions I keep returning to are practical ones: what an agent should remember, which decisions it may take on its own, and how people stay in control of systems that act for them.
Outcome
The findings become working tools and notes.