Running agents as a delivery team
Treating coding agents as a team with defined roles, a written backlog and CI that rejects work breaking the architecture — rather than as autocomplete.
I lead an engineering team, build things of my own, and keep a close eye on where software goes next.
I’m Twinkle Savani, an engineering manager based in Berlin. My background is backend and infrastructure — large-scale systems, and the unglamorous layers underneath them that simply have to be correct. Alongside that I build independent products, and spend much of my time working with AI coding agents to understand how software actually gets built now.
I grew up in India and live in Berlin. This site is the running record of that — the work, the notes, and some of the life around both.
Engineering. Leading backend and infrastructure work, and thinking about how much of an engineering manager’s job changes when the team’s output is no longer bounded by how fast people can type.
Building. The product I’m building is past the skeleton stage and moving story by story through the backlog. Most of the interesting decisions right now are product decisions, not technical ones.
Agents. Running Claude and Codex as an actual delivery team rather than as autocomplete: agents with defined roles, a written backlog, and CI that refuses work which breaks the architecture. Most of what I’m learning is about the guardrails, not the prompts.
Elsewhere. Reading about robotics and physical AI, and travelling when the calendar allows.
Treating coding agents as a team with defined roles, a written backlog and CI that rejects work breaking the architecture — rather than as autocomplete.
An independent consumer product, where most of the design work went into what it refuses to do rather than what it offers.
An autonomous SDLC orchestrator: take a GitHub issue, interrogate it, plan it, and open a tested pull request — with human approval kept explicit.
Process as the tool, guardrails instead of reminders, AI to raise the ceiling — and the part of the job that is still entirely about people.
A feed that cannot contain spam by construction does not need a spam policy. Moderation is a cost that grows with your users; a constraint is paid once.
If a product has to tell someone they weren't chosen, the design has already failed. Building the absence of a state turns out to be a real constraint.
A masters in Rajkot, an internship in Ahmedabad, a first job at TCS, then Mumbai, Bangalore, and eventually Berlin.
Base rather than backdrop — dancing, training, music, people, and a standing habit of leaving it regularly.
I also advise a small number of early-stage teams —what that looks like.