The tool I have quietly run my work week on is now open source.
Kiro Crew is a persistent agent workspace built on Kiro. I have been on the internal version since April, and it changed how I think about what AI at work actually is.
A while back I wrote about terminal productivity as a TPM and how far the command line can take a non-engineer. This is the next chapter: the terminal now runs itself.
I am a technical program manager, not a software engineer. My setup today: around a dozen scheduled jobs work on my behalf. They prepare my morning brief, watch releases and metrics, keep task trackers and reviews in sync, and speak up only when something needs me. The agent carries months of memory. My projects, my writing rules, the decisions I have already made. I never re-explain anything. It is not a chatbot I visit. It is a colleague already working when I arrive.
Under the hood it is one always-on gateway on my machine: a scheduler for the standing jobs, persistent memory that survives every restart, subagents for parallel work, and a lessons system that saves every correction I make so it never repeats a mistake. That last part matters more than any of it.
What surprised me was not the speed but the scope. The launch post from the three engineers who built it describes what happened inside Amazon: a side project became "something maintained by a community across many roles, not just engineers." In under six months it reached more than 39,000 builders, with nearly 500 contributors shipping 597 updates. I am one of those roles. I now prototype features, review architecture, and make one change land everywhere it needs to, from task boards to monthly reviews to launch announcements. Work that used to sit outside my role, or spread across five tools, now completes end to end.
Here is the honest part. The unlock was never the technology. It was making my own decision-making explicit enough to delegate. The months I spent teaching this system my rules, what good writing looks like, which data sources to trust, when to interrupt me, are the reason it performs.
Here is what that looks like. The agent built me a chart of adoption by role. Every number in it was correct. The chart was still wrong. One row was a worldwide rate and another was a single country, sitting in the same column, which made one team look like the worst performer when it wasn't. Nothing in the data flagged that. The rule was in my head, not the spreadsheet. So I wrote it down: one population per column, label the scope, exceptions get their own chart. It has not made that mistake since. That is the shift. I spend less time producing the analysis and more time writing down the standard the analysis has to meet.
The code is open source now. The judgment you encode in it will be yours.
Along the way I used it deeply enough to contribute to its Slack integration. My first code contribution to a platform I use every day.
Kiro Crew is out now. Whether you build software every day or have never opened a terminal, it meets you where you are. Start with one decision you make every week. Teach it that. Let it take the job from there.
https://kiro.dev/crew | https://github.com/kirodotdev/KiroCrew | https://twitch.tv/kirodotdev
Top comments (0)