Working with Pi and GPT-5.5
A personal, technical, and slightly critical note on why agentic coding finally feels enjoyable to steer.
I used to treat AI coding tools like a faster autocomplete box. Useful, sometimes impressive, but still mostly a place where I asked for snippets and then carried the real shape of the project in my own head.
Working with Pi and GPT-5.5 feels different. The enjoyable part is not that it magically knows everything. It does not. The enjoyable part is that the workflow starts to feel like steering a small technical team: one agent can inspect the repo, read the right files, use skills, run commands, make edits, and come back with something concrete enough to argue with.
That shift matters. When the tool has context and can act, the conversation changes from “write me code” to “help me move this system forward.” I can stay closer to product taste, architecture, and judgment while the agent handles more of the mechanical exploration.
But this is also where the rough edges show up. Agentic workflows are fragile. If context drifts, the agent can sound confident while looking in the wrong place. If a tool fails, the whole flow can become weirdly brittle. If I am vague, it may optimize for motion instead of correctness. The future feels close, but it still breaks in very ordinary ways.
That is why I enjoy Pi most when I do not treat it like autopilot. I treat it like a capable junior engineer with tools, memory, and initiative. I give it direction, challenge its assumptions, make it inspect the real code, and keep taste in the loop.
GPT-5.5 makes that loop feel faster and more alive. Pi gives the model a place to operate: files, terminal, project context, reusable skills, and a workflow that can cross from planning to implementation. Together, they make building feel less like prompting a chatbot and more like collaborating with an opinionated system that can actually touch the work.
I am still critical of it. I do not trust it blindly. But I genuinely enjoy working this way. The fun comes from the steering: knowing when to let the agent run, when to interrupt, when to ask better questions, and when to pull the work back into human judgment.