My first commit on GitHub is from October 2015, an Android app for a university lab. Today much of the code I ship isn’t typed by me at all. It’s written by agents I run, against specs I wrote, and it waits on approvals I give. One of my repositories has close to a thousand commits, and not one of them is mine.
This is how that happened, in roughly the order it happened.
Android, and a long quiet
From 2016 to 2019 my public repositories are Android: interview assignments, take-home tasks, small apps to learn a new library. Then my GitHub goes almost silent for four years.
Those were startup years. I worked with four startups at different stages, and the work lived in their repositories, not mine. What I took from them wasn’t a particular stack. It was the habit of turning a vague idea into something people can use, quickly, with whatever tools the team had.
Trying everything (2023)
In 2023 I started building for myself again, and it shows in the commit history: Supabase, Redwood, Medusa, Astro, Flutter, a Spotify clone. One to eight commits each. I wasn’t finishing things. I was learning what the modern stack felt like from the inside.
First contact with LLMs (2024)
My first AI code is from January 2024: two small projects mixing Formula 1 data with LangChain. Motorsport has been a recurring excuse to build things ever since: an F1 MCP server, a motorsport podcast generator, a fan-engagement idea list.
By the middle of 2024 I was keeping a prompt library and building an “agent studio”. In November I built World Journal, a pipeline that read RSS feeds, rewrote them with Claude and published them to a blog by itself.
It taught me something I keep relearning. Left alone, the posts converged. Within two weeks the blog was publishing near-identical pieces with titles like “Gaia’s Lament” and “Echoes of Earth”. Unattended generation doesn’t drift towards variety. It drifts towards its own favourite voice. Curation isn’t a finishing touch; it’s the product.
The platform I kept rebuilding (2024–2025)
From August 2024 to April 2025 I started the same agent platform five times under five names, and shipped none of them. I’ve written that story up separately in The agent platform I kept rebuilding. The short version: I kept fusing jobs that belong to different products, and it only worked once I split them.
MCP, and a challenge I didn’t finish (2025)
In March 2025 I wrote about ten MCP servers in ten weeks. One of them, the ERPNext MCP Server, became my most-used open-source work. I wrote about what maintaining it taught me in Maintaining an MCP server people actually use.
In May I started 30 Days 30 Agents: one agent a day for a month, in public. I got to day 22. The first week ran on schedule, then each day took a little longer than a day, and by early June it had stopped. The twenty-two that exist are real, and day 16 brought me back to Android for the first time in six years: an offline voice assistant running whisper.cpp on the phone.
Superprocess (2025 →)
In August 2025 I started building Superprocess, which I co-founded. It automates business processes that need judgment. AI agents, business rules, system integrations and human approvers run in one governed workflow, and anything material or uncertain pauses for a named owner. It was called Periscope until March 2026.
My personal repositories went quiet for most of the next six months, which tells you where the time went.
Running agents instead of building them (2026)
At the start of 2026 I stopped building an agent platform and started living with one: a small team of agents on a server, with names, memory, a shared board and a git identity of their own. That’s in Working with a team of agents.
It changed how I build everything else. Projects now start with a design spec and an implementation plan, and agents do most of the implementation. MakerLord, an assistant for people building electronics, went from spec to a live app in four days. CoWatch turns a codebase into a single interactive report. And at the end of 2025 I published AI2030, an 85,000-word novel about AI consciousness, written entirely by AI.
SuperJackfruitLabs (2026)
The experiments around running agents grew into SuperJackfruitLabs, my solo lab. It’s where I ask the same questions as at Superprocess without a customer waiting on the answer. It is now a small suite:
- AgentPod runs the machines and agent runtimes.
- Salaam keeps agent sessions alive and durable.
- SuperPipeline holds the work, with human approval gates.
- SuperMessage is the chat client for rooms with agents in them.
- SuperToolkit lets agents use credentials without holding them.
- SuperWitness records what agents did and how good it was.
- SuperLibrary keeps what they made, with its provenance.
- SuperMD is the Markdown editor I write in, and AgentNagar is a city for agents and people that’s still a prototype.
Most of these exist because the previous one ran into a problem. That’s the nicest thing I can say about building in public: the next thing to build is usually obvious.
The thread through all of it
Looking back, I’ve been asking one question since 2024 in different forms: how do you let software act on your behalf without losing control of it? My answer has settled into one rule, which I wrote up in People retain authority. Agents can do more and more of the work. People keep the decisions, and the system makes that structural rather than polite.
The rest is how I work:
- Docs first, then strict TDD. The spec is the source of truth.
- Honest READMEs. Say what doesn’t work yet.
- Bring your own everything. Your key, your model, your runtime. Self-hostable by default.
- People retain authority.
If any of this overlaps with what you’re working on, I’d like to hear about it: mail@rakeshgangwar.com.