In an agentic coding workflow, the best thing you can do is start at the decomposition.
That is where the real work moved. When an agent turns a clean, bounded unit into working code in minutes, the leverage is no longer in the implementation. It is in how you carved up the problem. A unit with a clear definition of done is exactly what an agent executes well. "Make checkout better" is where it flails.
Same Ritual, Different Payoff
That reframes t-shirt sizing. The S/M/L label was always a proxy for human effort, and agents collapsed most of that cost. But the decomposition underneath the sizing exercise is the whole game now.
Same ritual, different payoff: what used to produce an estimate now produces a spec.
The First Half Hour
The catch is that getting to the decomposition is slow. Most of the effort goes into assembling context before you can think clearly: pulling the ticket, reading the comment thread, checking the diff, chasing the linked issue. You spend the first half hour just gathering enough to have an opinion.
And that cost is not evenly distributed. Hand someone a codebase they did not write, with years of tribal knowledge baked into comment threads and closed tickets, and context-gathering is not a half-hour tax anymore. It is the whole obstacle. You cannot decompose what you do not understand, and understanding an unfamiliar system is mostly archaeology. Automating that dig is the difference between stalling for days and thinking on day one.
That is the part I wanted gone. TShirtize reads a Monday.com, Jira, or GitHub item and gathers the context for you: description, comments, diff scope, linked work. Then it hands you what decomposition actually needs. A definition of done, written as acceptance criteria you could hand off. The clarifying questions nobody asked yet. Spike flags where the unknowns are real. And the test tasks most estimates quietly skip. A size comes with it, but the size is the byproduct.
So you skip the gathering and start where the value is: breaking the work down well enough to hand it off.
Its Own Proof
TShirtize is its own proof. It is polished, it does a real job, and it came out of about 3 hours of actual work spread across 4 days. Not because I am fast. Because once the work was decomposed into clean, bounded units, the implementation was cheap and the agent handled it. I keep waiting for the hidden tax to show up, but the honest read is simpler: decompose well up front, and the quality comes cheap. That still surprises me.